Tag: open evaluation

Using performance-driven deep learning models to understand sensory cortex

In a new perspective piece in Nature Neuroscience, Yamins & Dicarlo (2016) discuss the emerging methodology and initial results in the literature of using deep neural nets with millions of parameters optimised for task performance to explain representations in sensory cortical areas. These are important developments. The authors explain the approach very well, also covering the historical progression toward it and its future potential.  Here are the key features of the approach as outlined by the authors.

(1) Complex models with multiple stages of nonlinear transformation from stimulus to response are used to explain high-level brain representations. The models are “stimulus computable” in the sense of fully specifying the computations from a physical description of the stimulus to the brain responses (avoiding the use of labels or judgments provided by humans).

(2) The models are neurobiologically plausible and “mappable”, in the sense that their components are thought to be implemented in specific locations in the brain. However, the models abstract from many biological details (e.g. spiking, in the reviewed studies).

(3) The parameters defining a model are specified by optimising the model’s performance at a task (e.g. object recognition). This is essential because deep models have millions of parameters, orders of magnitude too many to be constrained by the amounts of brain-activity data that can be acquired in a typical current study.

(4) Brain-activity data may additionally be used to define affine transformations of the model representations, so as to (a) fine-tune the model to explain the brain representations and/or (b) define the relationship between model units and measured brain responses in a particular individual.

(5) The resulting model is tested by evaluating the accuracy with which it predicts the representation of a set of stimuli not used in fitting the model. Prediction accuracy can be assessed at different levels of description:

  1. as the accuracy of prediction of a stimulus-response matrix,
  2. as the accuracy of prediction of a representational dissimilarity matrix, or
  3. as the accuracy of prediction of task-information decodability (i.e. are the decoding accuracies for a set of stimulus dichotomies correlated between model and neural population?).

A key insight is that the neural-predictive success of the models derives from combining constraints on architecture and function.

  • Architecture: Neuroanatomical and neurophysiological findings suggest (a) that units should minimally be able to compute linear combinations followed by static nonlinearities and (b) that their network architecture should be deep with rich multivariate representations at each stage. 
  • Function: Biological recognition performance and informal characterisations of high-level neuronal response properties suggest that the network should perform a transformation that retains high-level sensory information, but also emphasises behaviourally relevant categories and semantic dimensions. Large sets of labelled stimuli provide a constraint on the function to be computed in the form of a lookup table.

Bringing these constraints together has turned out to enable the identification of models that predict neural responses throughout the visual hierarchies better than any other currently available models. The models, thus, generalise not just to novel stimuli (Yamins et al. 2014; Khaligh-Razavi & Kriegeskorte 2014; Cadieu et al. 2014), but also from the constraints imposed on the mapping (e.g. mapping images to high-level categories) to intermediate-level representational stages (Güçlü & van Gerven 2015; Seibert et al. PP2016). Similar results are beginning to emerge for auditory representations.

The paper contains a useful future outlook, which is organised into sections considering improvements to each of the three components of the approach:

  • model architecture: How deep, what filter sizes, what nonlinearities? What pooling and local normalisation operations?
  • goal definition: What task performance is optimised to determine the parameters?
  • learning algorithm: Can learning algorithms more biologically plausible than backpropagation and potentially combining unsupervised and supervised learning be used?

In exploring alternative architectures, goals, and learning algorithms, we need to be guided by the known neurobiology and by the computational goals of the system (ultimately the organism’s survival and reproduction). The recent progress with neural networks in engineering provides the toolbox for combining neurobiologically plausible components and setting their parameters in a way that supports task performance. Alternative architectures, goals, and learning algorithms will be judged by their ability to predict neural representations of novel stimuli and biological behaviour.

The final section reflects on the fact that the feedfoward deep convolutional models currently very successful in this area only explain the feedforward component of sensory processing. Recurrent neural net models, which are also rapidly conquering increasingly complex tasks in engineering applications, promise to address these limitations of the initial studies using deep nets to explain sensory brain representations.

This perspective paper will be of interest to a broad audience of neuroscientists not themselves working with complex computational models, who are looking for a big-picture motivation of the approach and review of the most important findings. It will also be of interest to practitioners of this new approach, who will value the historical review and the careful motivation of each of the components of the methodology.

 

Deep net representational geometries become more similar to the ventral stream as performance is optimised

 

[R6I7]

 

Seibert, Yamins, Ardila, Hong, DiCarlo, and Gardner compared a deep convolutional neural network for visual object recognition to human ventral-stream representations as measured with fMRI (PP). The network was similar to the one described in Krizhevsky et al. (2012), the network that won the ImageNet competition that year with a large increase in performance compared to previous computer vision systems. The representations in the layers of the Krizhevsky deep net and similar models have been compared to human and monkey brain representations at different stages of the ventral stream previously (Yamins et al. 2013, Yamins et al. 2014; Khaligh-Razavi & Kriegeskorte 2014; Cadieu et al. 2014; Güçlü & van Gerven 2015). The present study is consistent with the previous results, generalises this line work to an interesting new set of test images, and investigates how the representational similarity of the model layers to the brain areas evolves as model performance is optimised. Results suggest that the optimisation of recognition performance increases representational similarity to visual areas, even for early and mid-level visual areas.

Model architecture: The convolutional network was inspired by that of Krizhevsky et al. (2012), using similar convolutional filter sizes, rectified linear units, the same pooling and local normalisation procedures, and data from ImageNet for training on 1000-class categorisation. However, the input images were downsampled to a substantially smaller size (120 x 120 pixels, instead of 224 x 224 pixels). Another major modification was that two intermediate fully connected layers (which contain most of the parameters in Krizhevsky et al.’s net) were omitted. This is reported to have no significant effect on recognition performance on an independent ImageNet test set.

Training and test stimuli: Like Krizhevsky et al., Seibert et al. trained the network by backpropagation to classify objects into 1000 categories. They used the very large ImageNet set of labelled images for model training and then presented the network and two human subjects with a different set of more controlled images: 1,785 grayscale images of 3D renderings of objects in many positions and views, superimposed to random natural backgrounds.

Representational similarity analysis: The authors compared the representational dissimilarity matrices (RDMs) between model layers and brain areas. They first randomly selected 1000 model features from a given layer, then reweighted these features, stretching and squeezing the representational space along its original axes, so as to maximise the RDM correlation between the model layer and the brain region. The maximisation of the RDM correlation was performed on the basis of 15 of the images for each of the 64 objects (different positions, views, and backgrounds). Using the fitted weights, they then re-estimated the model RDMs on the basis of the other 12 position-view combinations for the same 64 objects and computed the RDM correlation (Spearman) between model layer and brain region.

 

 

ScreenShot741

Detail of Figure 1 from the paper: Grayscale stimulus images were created by superimposing 3D models to natural backgrounds. The set strikes an interesting balance between naturalness and control. There were 8 objects from each of 8 categories (animals, boats, cars, chairs, faces, fruits, planes, tables) and each object was presented in 27 or 28 different combinations of position (including entirely nonoverlapping positions), view, and natural background image. For each of the 8 x 8 = 64 objects, they averaged response patterns to all the images that contained it, so as to compute 64 x 64-entry representational dissimilarity matrices (RDMs) using 1-Pearson correlation as the distance measure.

 

Related previous work: This work is closely related to recent papers by Yamins et al. (2013; 2014), Khaligh-Razavi et al. (2014), Cadieu et al. (2014), and Güçlü & van Gerven (2015). Yamins et al. showed that performance-optimised convolutional network models explain primate-IT neuronal recordings, with models performing better at object recognition also better explaining IT. Khaligh-Razavi et al. compared 37 computational model representations, including the layers of the Krizhevsky et al. (2012) model and a range of popular computer vision features, to human fMRI and monkey recording data (Kiani et al. 2007) and found that only the deep convolutional net, which was extensively trained to emphasize categorical divisions, could fully explain the IT data. They also showed that early visual cortex is well accounted for by earlier layers of the deep convolutional network (and by Gabor representations and other computer vision features). Cadieu et al. (2014) showed that among 6 different models, only Krizhevsky et al. (2012) and an even more powerful deep convolutional network by Zeiler & Fergus (2013) separate the categories in the representational space to a degree comparable to IT cortex. Güçlü & van Gerven (2015) investigated to what extent each layer of the model could explain the representations in each visual area of the ventral stream, finding rough correspondences between lower, intermediate, and higher model representations and early, mid-level, and higher ventral-stream regions, respectively.

 

How does the present work go beyond previous studies? The most striking novel contribution of this study is the characterisation of how representational similarity to visual areas develops as the neural net’s performance is optimised from a random initialisation. Unlike Yamins et al. (2014) and Cadieu et al. (2014), this study compares a convolutional network to the human ventral stream and, unlike Khaligh-Razavi & Kriegeskorte (2014), each image was presented in many positions and views and with many different backgrounds. The data is from only two subjects, but each subject underwent 9 sessions, so the total data set is substantial. The human fMRI data set is exciting in that it systematically varies category, exemplar, and accidental properties (position, view, background). However, the authors averaged across different images of each of the objects. I wonder if this data set has further potential for future analyses that don’t average across responses to different images.

Comparing many model representations to each of the areas of the visual system is a challenge requiring multiple studies. It’s great to see another study comparing the layers (including pooling layers and intermediate convolutional stages) alongside several control models (V1-like, V2-like, HMAX), which hadn’t been compared to deep convolutional networks before.

 

ScreenShot756

Figure 2 from the paper: Successive stages of the human ventral stream (V1, V2, V4, LOC) are best explained by successive layers of a deep convolutional neural net model. The representational geometry in V1 most resembles that of a lower and an intermediate layer of the network. The representational geometry in V2 most resembles that of an intermediate layer. And the representational geometries of V4 and LOC most resemble that of a higher layer of the network. Categories are reflected in clusters of response patterns in V4 and even more strongly in LOC. The same holds for higher layers of the network model.

 

Strengths

  • Model predictions of brain representational geometries are analysed as a function of model performance. This nicely demonstrates that it is not just the architecture, but performance optimisation that drives successful predictions of representations across all levels of the ventral stream.
  • Adds to the evidence that deep convolutional neural networks can explain the feedforward component of the stagewise representational transformations in the ventral visual stream.
  • Rich stimulus set of 1785 images that strikes an interesting balance between naturalism and control, independently varying objects and accidental properties.
  • Multiple data sets in each subject. This fMRI data set could in the future support tests of a wide variety of models.

 

Weaknesses

  • Statistical procedures are not clearly described and not fully justified. What type of generalisation does the crossvalidation scheme test for? What is being bootstrapped? Why are normal and independence assumptions relied on for inference, when bootstrapping the objects would enable straightforward tests that don’t require these assumptions?
  • The analysis is based on average response patterns across many different images for each object. This renders results more difficult to interpret.
  • Only two subjects.

Overall, this is a very nice study and a substantial contribution to the literature. However, the averaging across responses elicited by different images complicates the interpretation of the results and the statistical analyses need to be improved, better described, and fully justified – as detailed below. Although the overall results described in this review appear likely to hold up, I am not confident that the inferential results for particular model comparisons are reliable. (If concerns detailed below were substantively addressed, I would consider adjusting the reliability rating.)

 

Issues to consider addressing in a revision

(1) Can averaged response patterns elicited by different individual images be interpreted?

If we knew a priori that a region represents the objects with perfect invariance to position, view and background, then averaging across many images of the same objects that differ in these variables would make sense. However, we know that none of the regions is really invariant to position, view, and background, and gradually achieving some tolerance is one of the central computational challenges. The averaging will have differential effects in different regions as tolerance increases along the ventral stream. I don’t understand how to interpret the RDM for V1 given that it is based on averaged patterns. The object positions and backgrounds vary widely. Presumably different images of the same object are represented totally differently in V1. The averages should then form a tighter cluster of patterns (by factor 4 after averaging 16 images). Isn’t it puzzling then that the resulting RDM is still significantly correlated with the model? To explain this, do we have to assume that V1 actually represents the objects somewhat tolerantly (perhaps through feedback)? In a high-level representation tolerant to variation of accidental properties and sensitive to categorical differences, we expect the representations of the different images for a given object to be much more similar, so the averaging would have a smaller effect. All this confusion could be avoided by analysing patterns evoked by individual images. In addition, the emergence of tolerance across stages of processing could then be characterised.

 

(2) What type of generalisation does the crossvalidation scheme support?

Ideally, the crossvalidation should estimate the generalisation performance of the RDM prediction from the model for new images showing different objects. This is not the case here.

  • First, it appears that the brain data used for training (model weight fitting) and test (estimation of RDM correlation) are responses to the same set of images (all images). The weighting of the model features is estimated using a subset of 15 of the images for each object, and the RDM correlation between model and brain data assessed using 12 different images (different poses and backgrounds) of the same objects. This would seem to fall short of a test of generalisation to new images (even of the same objects) because all images are used (on the side of the brain responses) in the training procedure. Please clarify this issue.
  • Second, even if there was no overlap in the images used in training and test (on the side of either the model fitting or the brain data), the models are overfitted to the object set. Ideally, nonoverlapping sets of images of different objects should be used for training and testing. How about using a random subset of 4 of the objects in each category (32 in total) for fitting the weights and the other 32 objects for estimating the RDM correlation?

Overall, it seems unclear what type of generalisation these analyses test for. Let’s consider the issue of overfitting to the object set more closely. Currently, the weights w are fitted to 15 of the 27 images for a given object. In an idealised high-level representation invariant to the accidental variation, the two image sets will be identically represented. We expect the object representations to be in general position (no two on a point, no three on a line, no four on a plane and so on). Even if the 64 object representations were not at all clustered by category, but instead distributed randomly, we could linearly read out any categorical distinction and the decoder would generalise to the other 12 images. This is just to illustrate the expected effect of overfitting to the object set. In the present study, weights were fitted to predict RDMs not to discriminate categories. Fitting the 1000 scaling parameters to explain an RDM with 64*(64-1)/2 = about 2K dissimilarities should enable us to fit any RDM quite precisely. I would not be surprised if a noncategorical representation could fit a clearly categorical representation (block-diagonal RDM) in this context. The test-set correlations would then really just be a measure of the replicability of the brain RDMs – rather than a measure of the fit of the model. Regularisation might help ensure that different models are still distinguishable, but it also further complicates interpretation (see below).

Since higher regions are more tolerant, the training and test images are more similarly represented in these regions, and so we would expect greater positive overfitting bias on the estimated RDM correlation for higher regions regardless of the model. It is reassuring that the models still perform differently in LOC. However, the overfitting to the object set complicates interpretation.

The category decoding performance measure is similarly compromised by averaging across different images. Decoding performance as well (if I understood correctly) was tested by averaging different images for each object and training and testing on the same set of objects with different particular images in the test set. So the test is not a test of generalisation to different objects but to different images of the same objects. Again any representation uniquely representing each object (and having at least as many dimensions as the number of objects in both classes combined minus one, which is the case here) will appear to support linear category decoding, even if the distributions in representational space corresponding to the two categories (including the entire populations of objects they comprise) were not at all linearly separable and across-object generalisation performance were at chance level.

 

(3) Clarify the bootstrapping procedure used in model comparison

The first 6 times the term bootstrap is used, it is entirely unclear what entities are being resampled with replacement. The sampling of 1000 model units is explained in this context, and suggests that this is the resampling with replacement referred to as bootstrapping. Only on page 15 it says: “Our approach bootstraps over independent stimulus samples”. I’m not sure what multiple independent stimulus samples are meant here. Are the objects (averaged across images) resampled? Or are the images resampled? (The latter would necessitate re-estimating the object-average voxel responses to each object for each bootstrap sample.)

 

(4) Clarify and justify the test used for model comparison

The methods section states:

“Using the bootstrapping above, we computed p-values testing if Layer A better explained visual area X’s RDM than Layer B”

This suggests that a bootstrap test was used to compare models with respect to their RDM prediction performance. But then the model comparison test is described as follows:

“We use Fisher’s r-to-z transformation using Steiger (1980)’s approach to compute p-values for difference in correlation values (Lee and Preacher, 2013). The approach tests for equality of two correlation values from the same sample where one variable is held in common between the two coefficients (in our case, an RDM of a given visual area).”

The Steiger (1980) method for comparing two dependent correlations assumes that the elements of the correlated vectors are sampled independently. As you acknowledge, this is violated for dissimilarities in an RDM. But why then is the Steiger method appropriate? You mention bootstrapping, but don’t explain how your bootstrapping procedure interacts with the Steiger method. Kriegeskorte et al. (2008) and Nili et al. (2014) describe a bootstrapping approach to RDM model comparison that takes the dependencies between dissimilarities into account and does not rely on the Steiger method. Ideally, the objects, not the images, should be resampled with replacement, to simulate variation across objects (not across images of the same objects) and to avoid re-estimation of object-average patterns. Finally, it would be good to use model-comparative inference to support the improvement the RDM explanation of ventral stream regions as performance is optimised by training with backpropagation.

 

(5) Reconsider the regularisation used in feature-weight fitting

The one-iteration optimisation (motivated as a variant of early stopping) is a very ad-hoc choice of regulariser. I have no idea what prior is implicit to this method. However, this implicit prior is part of the model you are testing and affects the model comparison results. It is even a key component of the model because you are fitting so many parameters that different models might not be distinguishable without this prior.

 

(6) Show full inferential results with correction for multiple testing

It would be great if the figures showed which RDM correlations are significant and which pairs are significantly different. In addition, it would be good to account for multiple testing. Nonparametric methods for testing and comparing RDM correlations are described in Nili et al. (2014).

 

(7) Show noise ceilings

It would be good to see whether the model layers fully or only partially explain the explainable component of the variance in the RDMs. This could yield the insight that the model does fall short given the present study’s data set. It would be interesting then to learn how it falls short and this would motivate future changes to the model. Alternatively, if the model reaches the noise ceiling, we would learn that we need to get better or more data to find out how the model still falls short. The methods section suggests that a noise ceiling was estimated for Figure 6, stating:

“To avoid the problem of finding linear re-weightings using smaller sub-sets of our data, we instead computed noise ceilings and percent explained variance values (Figure 6) without using the weighting procedure described above. Noise ceilings for each visual area were computed by splitting the runs of our data into two non-overlapping groups. With each group, we estimated stimulus responses (beta weights) using the procedure described above (see the Image responses section) and computed object-averaged RDMs for each visual area. We used the correlation between the RDM from each of the two groups as our noise ceiling for percent explained variance estimates (Figure 6).”

However, Figure 6 and its legend don’t mention a noise ceiling. What is the noise ceiling in these analyses? Figure 2 would also benefit from noise ceilings for each of the brain areas. In addition, the split-half correlation should underestimate the noise ceiling because half the data is used and both RDMs are affected by noise. The noise ceiling computation should instead give an estimate of the expected performance of a noiseless true model or upper and lower bounds on this performance (Nili et al. 2014, Khaligh-Razavi & Kriegeskorte 2015).

 

(8) What is the function of the two branches of the model?

Clarify the function of the two branches of the model in the legend of Figure 5 and in the methods section. A single GPU was used for training here. Did this serve to keep the architecture consistent with Krizhevsky et al. (2012)?

 

(9) Why are the ROIs so big?

As far as I remember, a normal size for LO or FFA is below 1 ml. LO1, LO2, and FFA have 234, 299, and 292 voxels (pooled across two subjects), corresponding to 3 to 4 ml on average across subjects (given that voxels were 3 mm isotropic).

 

(10) Add a colour legend to Figure 4

This would help the reader quickly understand the meaning of the lines without having to refer to the text description.

 

 

The four pillars of open science

An open review of Gorgolewski & Poldrack (PP2016)

the 4 pillars of open science.png

The four pillars of open science are open data, open code, open papers (open access), and open reviews (open evaluation). A practical guide to the first three of these is provided by Gorgolewski & Poldrack (PP2016). In this open review, I suggest a major revision in which the authors add treatment of the essential fourth pillar: open review. Image: The Porch of the Caryatids (Porch of the Maidens) of the ancient Greek temple Erechtheion on the north side of the Acropolis of Athens.

 

Open science is a major buzz word. Is all the talk about it just hype? Or is there a substantial vision that has a chance of becoming a reality? Many of us feel that science can be made more efficient, more reliable, and more creative through a more open flow of information within the scientific community and beyond. The internet provides the technological basis for implementing open science. However, making real progress with this positive vision requires us to reinvent much of our culture and technology. We should not expect this to be easy or quick. It might take a decade or two. However, the arguments for openness are compelling and open science will prevail eventually.

The major barriers to progress are not technological, but psychological, cultural, and political: individual habits, institutional inertia, unhealthy incentives, and vested interests. The biggest challenge is the fact that the present way of doing science does work (albeit suboptimally) and our vision for open science has not merely not yet been implemented, but has yet to be fully conceived. We will need to find ways to gradually evolve our individual workflows and our scientific culture.

Gorgolewski & Poldrack (PP2016) offer a brief practical guide to open science for researchers in brain imaging. I was expecting a commentary reiterating the arguments for open science most of us have heard before. However, the paper instead makes good on its promise to provide a practical guide for brain imaging and it contains many pointers that I will share with my lab and likely refer to in the future.

The paper discusses open data, open code, and open publications – describing tools and standards that can help make science more transparent and efficient. My main criticism is that it leaves out what I think of as a fourth essential pillar of open science: open peer review. Below I first summarise some of the main points and pointers to resources that I took from the paper. Along the way, I add some further points overlooked in the paper that I feel deserve consideration. In the final section, I address the fourth pillar: open review. In the spirit of a practical guide, I suggest what each of us can easily do now to help open up the review process.

 

1 Open data

  • Open-data papers more cited, more correct: If data for a paper are published, the community can reanalyse the data to confirm results and to address additional questions. Papers with open data are cited more (Piwowar et al. 2007, Piwowar & Vision 2013) and tend to make more correct use of statistics (Wicherts et al. 2011).
  • Participant consent: Deidentified data can be freely shared without consent from the participants in the US. However, rules differ in other countries. Ideally, participants should consent to their data being shared. Template text for consent forms is offered by the authors.
  • Data description: The Brain Imaging Data Structure (BIDS) (Gorgolewski et al. 2015) provides a standard (evolved from the authors’ OpenfMRI project; Poldrack et al. 2013) for file naming and folder organisation, using file formats such as NifTI, TSV and JSON.
  • Field-specific brain-imaging data repositories: Two repositories accept brain imaging data from any researcher: FCP/INDI (for resting state fMRI only) and OpenfMRI (for any datasets that includes MRI data).
  • Field-general repositories: Field-specific repositories like those mentioned help standardise sharing for particular types of data. If the formats offered are not appropriate for the data to be shared, field-general repositories, including FigShare, Dryad, or DataVerse can be used.
  • Data papers: A data paper is a paper that focusses on the description of a particular data set that is publicly accessible. This helps create incentives for ambitious data acquisitions and to enable researchers to specialise in data acquisition. Journals publishing data papers include: Scientific Data, Gigascience, Data in Brief, F1000Research, Neuroinformatics, and Frontiers in Neuroscience.
  • Processed-data sharing: It can be useful to share intermediate or final results of data analysis. With the initial (and often somewhat more standardised) steps of data processing out of the way, processed data are often much smaller in volume and more immediately amenable to further analyses by others. Statistical brain-imaging maps can be shared via the authors’ NeuroVault.org website.

 

2 Open code

  • Code sharing for transparency and reuse: Data-analysis details are complex in brain imaging, often specific to a particular study, and seldom fully defined in the methods section. Sharing code is the only realistic way of fully defining how the data have been analysed and enabling others to check the correctness of the code and effects of adjustments. In addition, the code can be used as a starting point for the development of further analyses.
  • Your code is good enough to share: A barrier to sharing is the perception among authors that their code might not be good enough. It might be incompletely documented, suboptimal, or even contain errors. Until the field finds ways to incentivise greater investment in code development and documentation for sharing, it is important to lower the barriers to sharing. Sharing imperfect code is preferable to not sharing code (Barnes 2010).
  • Sharing does not imply provision of user support: Sharing one’s code does not imply that one will be available to provide support to users. Websites like org can help users ask and answer questions independently (or with only occasional involvement) of the authors.
  • Version Control System (VCS) essential to code sharing: VCS software enables maintenance of complex code bases with multiple programmers and versions, including the ability to merge independent developments, revert to previous versions when a change causes errors, and to share code among collaborators or publicly. An excellent, freely accessible, widely used, web-based VCS platform is com, introduced in Blischak et al. (2016).
  • Literate programming combines code and results and text narrative: Scripted automatic analyses have the advantage of automaticity and reproducibility (Cusack et al. 2014), compared to point-and-click analysis in an application with a graphical user interface. However, the latter enables more interactive interrogation of the data. Literate programming (Knuth 1992) attempts to make coding more interactive and provides a full and integrated record of the code, results, and text explanations. This provides a fully computationally transparent presentation of results, makes the code accessible to oneself later in time, and to collaborators and third parties, with whom literate programs can be shared (e.g. via GitHub). Software supporting this includes: Jupyter (for R, Python and Julia), R Markdown (for R) and matlabweb (for MATLAB).

 

3 Open papers

  • Open notebook science: Open science is about enhancing the bandwidth and reducing the latency in our communication network. This means sharing more and at earlier stages, not only our data and code, but ultimately also our day-to-day incremental progress. This is called open notebook science and has been explored, by Cameron Neylon and Michael Nielson among others. Gorgolewski & Poldrack don’t comment on this beautiful vision for an entirely different workflow and culture at all. Perhaps open notebook science is too far in the future? However, some are already practicing it. Surely, we should start exploring it in theory and considering what aspects of open notebook science we can integrate into our workflow. It would be great to have some pointers to practices and tools that help us move in this direction.
  • The scientific paper remains a critical component of scientific communication: Data and code sharing are essential, but will not replace communication through permanently citable scientific papers that link (increasingly accessible) data through analyses to novel insights and relate these insights to the literature.
  • Papers should simultaneously achieve clarity and transparency: The conceptual clarity of the argument leading to an insight is often at a tension with the transparency of all the methodological details. Ideally, a paper will achieve both clarity and transparency, providing multiple levels of description: a main narrative that abstracts from the details, more detailed descriptions in the methods section, additional detail in the supplementary information, and full detail in the links to the open data and code, which together enable exact reproduction of the results in the figures. This is an ideal to aspire to. I wonder if any paper in our field has fully achieved it. If there is one, it should surely be cited.
  • Open access: Papers need to be openly accessible, so their insights can have the greatest positive impact on science and society. This is really a no brainer. The internet has greatly lowered the cost of publication, but the publishing industry has found ways to charge higher prices through a combination of paywalls and unreasonable open-access charges. I would add that every journal contains unique content, so the publishing industry runs hundreds of thousands of little monopolies – safe from competition. Many funding bodies require that studies they funded be published with open access. We need political initiatives that simply require all publicly funded research to be publicly accessible. In addition, we need publicly funded publication platforms that provide cost-effective alternatives to private publishing companies for editorial boards that run journals. Many journals are currently run by scientists whose salaries are funded by academic institutions and the public, but whose editorial work contributes to the profits of private publishers. In historical retrospect, future generations will marvel at the genius of an industry that managed for decades to employ a community without payment, take the fruits of their labour, and sell them back to that very community at exorbitant prices – or perhaps they will just note the idiocy of that community for playing along with this racket.
  • Preprint servers provide open access for free: Preprint servers like bioRxiv and arXiv host papers before and after peer review. Publishing each paper on a preprint server ensures immediate and permanent open access.
  • Preprints have digital object identifiers (DOIs) and are citable: Unlike blog posts and other more fleeting forms of publication, preprints can thus be cited with assurance of permanent accessibility. In my lab, we cite preprints we believe to be of high quality even before peer review.
  • Preprint posting enables community feedback and can help establish precedence: If a paper is accessible before it is finalised the community can respond to it and help catch errors and improve the final version. In addition, it can help the authors establish the precedence of their work. I would add that this potential advantage will be weighed against the risk of getting scooped by a competitor who benefits from the preprint and is first to publish a peer-reviewed journal version. Incentives are shifting and will encourage earlier and earlier posting. In my lab, we typically post at the time of initial submission. At this point getting scooped is unlikely, and the benefits of getting earlier feedback, catching errors, and bringing the work to the attention of the community outweighs any risks of early posting.
  • Almost all journals support the posting of preprints: Although this is not widely known in the brain imaging and neuroscience communities, almost all major journals (including Nature, Science, Nature Neuroscience and most others) have preprint policies supportive of posting preprints. Gorgolewski & Poldrack note that they “are not aware of any neuroscience journals that do not allow authors to deposit preprints before submission, although some journals such as Neuron and Current Biology consider each submission independently and thus one should contact the editor prior to submission.” I would add that this reflects the fact that preprints are also advantageous to journals: They help catch errors and get the reception process and citation of the paper going earlier, boosting citations in the two-year window that matters for a journal’s impact factor.

 

4 Open reviews

The fourth pillar of open science is the open evaluation (OE, i.e. open peer review and rating) of scientific papers. This pillar is entirely overlooked in the present version of the Gorgolewski & Poldrack’s commentary. However, peer review is an essential component of communication in science. Peer review is the process by which we prioritise the literature, guiding each field’s attention, and steering scientific progress. Like other components of science, peer review is currently compromised by a lack of transparency, by inefficiency of information flow, and by unhealthy incentives. The movement for opening the peer review process is growing.

In traditional peer review, we judge anonymously, making inherently subjective decisions that decide about the publication of our competitors’ work, under a cloak of secrecy and without ever having to answer for our judgments. It is easy to see that this does not provide ideal incentives for objectivity and constructive criticism. We’ve inherited secret peer review from the pre-internet age (when perhaps it made sense). Now we need to overcome this dysfunctional system. However, we’ve grown used to it and may be somewhat comfortable with it.

Transparent review means (1) that reviews are public communications and (2) that many of them are signed by their authors. Anonymous reviewing must remain an option, to enable scientists to avoid social consequences of negative judgments in certain scenarios. However, if our judgment is sound and constructively communicated, we should be able to stand by it. Just like in other domains, transparency is the antidote to corruption. Self-serving arguments won’t fly in open reviewing, and even less so when the review is signed. Signing adds weight to a review. The reviewer’s reputation is on the line, creating a strong incentive to be objective, to avoid any impression of self-serving judgment, and to attempt to be on the right side of history in one’s judgment of another scientist’s work. Signing also enables the reviewer to take credit for the hard work of reviewing.

The arguments for OE and a synopsis of 18 visions for how OE might be implemented are given in Kriegeskorte, Walther & Deca (2012). As for other components of open science, the primary obstacles to more open practices are not technological, but psychological, cultural, and political. Important journals like eLife and those of the PLoS family are experimenting with steps toward opening the review process. New journals including, the Winnower, ScienceOpen, and F1000 Research already rely on postpublication peer review.

We don’t have to wait for journals to lead us. We have all the tools to reinvent the culture of peer review. The question is whether we can handle the challenges this poses. Here, in the spirit of Gorgolewki & Poldrack’s practical guide, are some ways that we can make progress toward OE now by doing things a little differently.

  • Sign peer reviews you author: Signing our reviews is a major step out of the dark ages of peer review. It’s easier said than done. How can we be as critical as we sometimes have to be and stand by our judgment? We can focus first on the strengths of a paper, then communicate all our critical arguments in a constructive manner. Some people feel that we must sign either all or none of our reviews. I think that position is unwise. It discourages beginning to sign and thus de facto cements the status quo. In addition, there are cases where the option to remain anonymous is needed, and as long as this option exists we cannot enforce signing anyway. What we can do is take anonymous comments with a grain of salt and give greater credence to signed reviews. It is better to sign sometimes than never. When I started to sign my reviews, I initially reserved the right to anonymity for myself. After all this was a unilateral act of openness; most of my peers do not sign their reviews. However, after a while, I decided to sign all of my reviews, including negative ones.
  • Openly review papers that have preprints: When we read important papers as preprints, let’s consider reviewing them openly. This can simultaneously serve our own and our collective thought process: an open notebook distilling the meaning of a paper, why its claims might or might not be reliable, how it relates to the literature, and what future steps it suggests. I use a blog. Alternatively or additionally, we can use PubMed Commons or PubPeer.
  • Make the reviews you write for journals open: When we are invited to do a review, we can check if the paper has been posted as a preprint. If not, we can contact the authors, asking them to consider posting. At the time of initial submission, the benefits tend to outweigh the risks of posting, so many authors will be open to this. Preprint posting is essential to open review. If a preprint is available, we can openly review it immediately and make the same review available to the journal to contribute to their decision process.
  • Reinvent peer review: What is an open review? For example, what is this thing you’re reading? A blog post? A peer review? Open notes on the essential points I would like to remember from the paper with my own ideas interwoven? All of the above. Ideally, an open review helps the reviewer, the authors, and the community think – by explaining the meaning of a paper in the context of the literature, judging the reliability of its claims, and suggesting future improvements. As we begin to review openly, we are reinventing peer review and the evaluation of scientific papers.
  • Invent peer rating: Eventually we will need quantitative measures evaluating papers. These should not be based on buzz and usage statistics, but reflect the careful judgement of peers who are experts in the field, have considered the paper in detail, and ideally stand by their judgment. Quantitative judgments can be captured in ratings. Multidimensional peer ratings can be used to build a plurality of paper evaluation functions (Kriegeskorte 2012) that prioritise the literature from different perspectives. We need to invent suitable rating systems. For primary research papers, I use single-digit ratings on multiple scales including reliability, importance, and novelty, using capital letters to indicate the scale in the following format: [R7I5].

 

Errors are normal

As we open our science and share more of it with the community, we run the risk of revealing more of our errors. From an idealistic perspective that’s a good thing, enabling us learn more efficiently as individuals and as a community. However, in the current game of high-impact biomedical science there is an implicit pretense that major errors are unlikely. This is the reason why, in the rare case that a major error is revealed despite our lack of transparent practices, the current culture requires that everyone act surprised and the author be humiliated. Open science will teach us to drop these pretenses. We need to learn to own our mistakes (Marder 2015) and to be protective of others when errors are revealed. Opening science is an exciting creative challenge at many levels. It’s about reinventing our culture to optimise our collective cognitive process. What could be more important or glamorous?

 

Additional suggestions for improvements in revision

  • A major relevant development regarding open science in the brain imaging community is the OHBM’s Committee on Best Practices in Data Analysis and Sharing (COBIDAS), of which author Russ Poldrack and I are members. COBIDAS is attempting to define recommended practices for the neuroimaging community and has begun a broad dialogue with the community of researchers (see weblink above). It would be good to explain how COBIDAS fits in with the other developments.
  • About a third of the cited papers are by the authors. This illustrates their substantial contribution and expertise in this field. I found all these papers worthy of citation in this context. However, I wonder if other groups that have made important contributions to this field should be more broadly cited. I haven’t followed this literature closely enough to give specific suggestions, but perhaps it’s worth considering whether references should be added to important work by others.
  • As for the papers, the authors are directly involved in most of the cited web resources OpenfMRI, NeuroVault, NeuroStars.org. This is absolutely wonderful, and it might just be that there is not much else out there. Perhaps readers of this open review can leave pointers in the comments in case they are aware of other relevant resources. I would share these with the authors, so they can consider whether to include them in revision.
  • Can the practical pointers be distilled into a table or figure that summarises the essentials? This would be a useful thing to print out and post next to our screens.
  • “more than fair” -> “only fair”

 

Disclosures

I have the following relationships with the authors.

relationship number of authors
acquainted 2
collaborated on committee 1
collaborated on scientific project 0

 

References

Barnes N (2010) Publish your computer code: it is good enough. Nature. 467: 753. doi: 10.1038/467753a

Blischak JD, Davenport ER, Wilson G. (2016) A Quick Introduction to Version Control with Git and GitHub. PLoS Comput Biol. 12: e1004668. doi: 10.1371/journal.pcbi.1004668

Cusack R, Vicente-Grabovetsky A, Mitchell DJ, Wild CJ, Auer T, Linke AC, et al. (2014) Automatic analysis (aa): efficient neuroimaging workflows and parallel processing using Matlab and XML. Front Neuroinform. 2014;8: 90. doi: 10.3389/fninf.2014.00090

Gorgolewski KJ, Auer T, Calhoun VD, Cameron Craddock R, Das S, Duff EP, et al. (2015) The Brain Imaging Data Structure: a standard for organizing and describing outputs of neuroimaging experiments [Internet]. bioRxiv. 2015. p. 034561. doi: 10.1101/034561

Gorgolewski KJ, Varoquaux G, Rivera G, Schwarz Y, Ghosh SS, Maumet C, et al. (2015) NeuroVault.org: a webbased repository for collecting and sharing unthresholded statistical maps of the human brain. Front Neuroinform. Frontiers. 9. doi: 10.3389/fninf.2015.00008

Knuth DE (1992) Literate programming. CSLI Lecture Notes, Stanford, CA: Center for the Study of Language and Information (CSLI).

Kriegeskorte N, Walther A, Deca D (2012) An emerging consensus for open evaluation: 18 visions for the future of scientific publishing Front. Comput. Neurosci http://dx.doi.org/10.3389/fncom.2012.00094

Kriegeskorte N (2012) Open evaluation: a vision for entirely transparent post-publication peer review and rating for science. Front. Comput. Neurosci., 17 http://dx.doi.org/10.3389/fncom.2012.00079

Marder E (2015) Living Science: Owning your mistakes DOI: http://dx.doi.org/10.7554/eLife.11628 eLife 2015;4:e11628

Piwowar HA, Day RS, Fridsma DB (2007) Sharing detailed research data is associated with increased citation rate. PLoS One. 2007;2: e308. doi: 10.1371/journal.pone.0000308

Piwowar HA, Vision TJ (2013) Data reuse and the open data citation advantage. PeerJ. 1: e175. doi: 10.7717/peerj.175

Poldrack RA, Barch DM, Mitchell JP, Wager TD, Wagner AD, Devlin JT, et al. (2013) Toward open sharing of taskbased fMRI data: the OpenfMRI project. Front Neuroinform. 2013;7: 1–12. doi: 10.3389/fninf.2013.00012

Wicherts JM, Bakker M, Molenaar D (2011) Willingness to Share Research Data Is Related to the Strength of the Evidence and the Quality of Reporting of Statistical Results. Tractenberg RE, editor. PLoS One. 6: e26828. doi: 10.1371/journal.pone.0026828

 

 

 

 

 

 

 

 

Imagining and seeing objects elicits consistent category-average activity patterns in the ventral stream

[R8I7]

Horikawa and Kamitani report results of a conceptually beautiful and technically sophisticated study decoding the category of imagined objects. They trained linear models to decode visual image features from fMRI voxel patterns. The visual features are computed from images by computational models including GIST and the AlexNet deep convolutional neural net. AlexNet provides features spanning the range from visual to semantic. A subject is then scanned while imagining images from a novel object category (not used in training the fMRI decoder). The decoder is used to predict the computational-model representation for the imagined category (averaged across exemplars of that category). This predicted model representation is then compared to the actual model representation for many categories, including the imagined one. The model representation predicted from fMRI during imagery is shown to be significantly more similar to the model representation of images from the imagined category than to the model representation of images from other categories.

ScreenShot685

Figure from Horikawa & Kamitani (2015)

The methods are sophisticated and will give experts much to think about and draw from in developing better decoders. Comprehensive supplementary analyses, which I did not have time to fully review, complement and extend the thorough analyses provided. This is a great study. As usual in our field, a difficult question is what exactly it means for brain computational theory.

A few results that might speak to the computational mechanism of the ventral stream are as follows.

When predicting computational features of *single images* (which was only done for seen, not for imagined objects):

  • Lower layers of AlexNet are better predicted from voxels in lower ventral-stream areas.
  • Higher layers of AlexNet are better predicted from voxels in higher ventral-stream areas.
  • GIST features are best predicted from V1-3, but also significantly from higher areas.

This is consistent with the recent findings (Yamins, Khaligh-Razavi, Cadieu, Guclu) showing that deep convolutional neural nets explain lower- and higher-level ventral-stream areas with a rough correspondence of lower model layers to lower brain areas and higher model layers to higher brain areas. It is also consistent with previous findings that GIST, like many visual feature models, explains significant representational variance even in the higher ventral-stream representation (Khaligh-Razavi, Rice), but does not reach the noise ceiling (indicating that a data set is fully explained), as deep neural net models do (Khaligh-Razavi).

When predicting *category-averages* of computational features (which was done for seen and imagined objects):

  • Higher-level visual areas better predict features in all layers of AlexNet.
  • Higher layers of AlexNet are better predicted from voxels in all visual areas.

This is confusing, until we remember that it is category averages that are being predicted. Category averaging will retain a major portion of the representational variance of category-sensitive higher-level representations, while reducing the representational variance of low-level representations that are less related to categories. This may boost both predictions from category-related visual areas, as well as predictions of category-related model features.

Subjects imagined many different images from a given category in an experimental block during fMRI. The category-average imagery activity of the voxels was then used to predict the corresponding category-averages of the computational-model features. As expected, category-average computational-feature prediction is worse for mental imagery than for perception. The pattern across visual areas and AlexNet layers is similar for imagery and perception, with higher predictions resulting when the predicting visual area is category-related and when the predicted model feature is category-related. However, V1 and V2 did not consistently enable imagery decoding into the format of any of the layers of AlexNet. Interestingly, computational features more related to categories were better decodable. This supports the view that higher ventral-stream features might be optimised to emphasise categorical divisions (cf Jozwik et al. 2015).

 

Suggested improvements

(1) Clarify any evidence about the representational format in which the imagined content is represented. The authors’ model predicts both visual and semantic features of imagined object categories. This suggests that imagery involves both semantic and visual representations. However, the evidence for lower- or even mid-level visual representation of imagined objects is not very compelling here, because the imagery was not restricted to particular images. Instead the category-average imagery activity was measured. Each category is, of course, associated with particular visual features to some extent. We therefore expect to be able to predict category-average visual features from category-average voxel patterns better than chance. A strong claim that imagery paints low-level visual features into early visual representations would require imagery of particular images within each category. For relevant evidence, see Naselaris et al. (2015).

(2) Go beyond the decoding spin: what do we learn about computations in the ventral stream? Being able to decode brain representations is cool because it demonstrates unambiguously that a certain kind of information is present in a brain region. It’s even cooler to be able to decode into an open space of features or categories and to decode internally generated representations as done here. Nevertheless, the approach of decoding is also scientifically limiting. From the present version of the paper, the message I take is summarised in the title of the review: “Imagining and seeing objects elicits consistent category-average activity patterns in the ventral stream”. This has been shown previously (e.g. Stokes, Lee), but is greatly generalised here and is a finding so important that it is good to have it replicated and generalised in multiple studies. The reason why I can’t currently take a stronger computational claim from the paper is that we already know that category-related activity patterns cluster hierarchically in the ventral stream (Kriegeskorte et al. 2008) and may be continuously and smoothly related to a semantic space (Mitchell et al. 2008; Huth et al. 2012). In the context of these two pieces of knowledge, consistent category-average activity for perception and imagery is all that is needed to explain the present findings of decodability of novel imagined categories. The challenge to the authors: Can you test specific computational hypotheses and show something more on the basis of this impressive experiment? The semantic space analysis goes in this direction, but did not appear to me to support totally novel theoretical conclusions.

(3) Why decode computational features? Decoding of imagined content could be achieved either by predicting measured activity patterns from model representations of the stimuli (e.g. Kay et al. 2008) or by predicting model representations  from measured activity patterns (the present approach). The former approach is motivated by the idea that the model should predict the data and lends itself to comparing multiple models, thus contributing to computational theory. We will see below that the latter approach (chosen here) is less well suited to comparing alternative computational models. Why did Horikawa & Kamitani choose this approach? One argument might be that there are many model features and predicting the smaller number of voxels from these many features requires strong prior assumptions (implicit to regularisation), which might be questionable. The reverse prediction from voxels to features requires estimating the same total number of weights (# voxels * # model features), but each univariate linear model predicting a feature only has # voxels (i.e. typically fewer than # features) weights. Is this why you preferred this approach? Does it outperform the voxel-RF modelling approach of Kay et al. (2008) for decoding?

An even more important question is what we can learn about brain computations from feature decoding. If V4, say, perfectly predicted CNN1, this would suggest that V4 contains features similar to those in CNN1. However, it might additionally contain more complex features unrelated to CNN1. CNN1 predictability from V4, thus, would not imply that CNN1 can account for V4. Another example: CNN8 and GIST features are similarly predictable from voxel data across brain areas, and most predictable from V4 voxels. Does this mean GIST is as good a model as CNN8 for explaining the computational mechanism of the ventral stream? No. Even if the ventral-stream voxels perfectly predicted GIST, this would not imply that GIST perfectly predicts the ventral-stream voxels.

The important theoretical question is what computational mechanism gives rise to the representation in each area. For the human inferior temporal cortex, Khaligh-Razavi & Kriegeskorte (2015) showed that both GIST and the CNN representation explain significant variance. However, the GIST representation leaves a large portion of the explainable variance unexplained, whereas the CNN fully explains the explainable variance.

(4) Further explore the nature of the semantic space. To understand what drives the decoding of imagined categories, it would be helpful to see the performance of simpler analyses. Following Mitchell et al. (2008), one could use a text-corpus based semantic embedding to represent each of the categories. Decoding into this semantic embedding would similarly enable novel seen and imagined test categories (not used in training) to be decoded. It would be interesting, then, to successively reduce the dimensionality of the semantic embedding to estimate the complexity of the semantic space underlying the decoding. Alternatively, the authors’ WordNet distance could be used for decoding.

(5) Clarify that category-average patterns were used. The terms “image-based information” and “object-based information” are not ideal. By “image-based”, you are referring to a low-level visual representation and by “object-based”, to a categorical representation. Similarly, in many places where you say “objects” (as in “decoding objects”) it would be clearer to say “object categories”. Use clearer language throughout to clarify when it was category-average patterns that were used for prediction (brain representations) and that were predicted (model representations). This concerns the text and the figures. For example, the title of Fig. 4 should be: “Object-category-average feature decoding”. If this detracts the casual reader too lazy to even read the legends too much, at least the text of the legend should clearly state that category-average brain activity patterns are used to predict category-average model features.

(6) What are the assumptions implicit to sparse linear regression and is this approach optimal? L2 regularisation would spread the weights out over more voxels and might benefit from averaging out the noise component. Please comment on this choice and on any alternative performance results you may have.

 

Minor points

(7) The work is related to Mitchell et al. (2008), who predicted semantic semantic brain representations of novel stimuli using a semantic space model. This paper should be cited.

(8) “These studies showed a high representational similarity between the top layer of a convolutional neural network and visual cortical activity in the inferior temporal (IT) cortex of humans [24,25] and non-human primates [22,23].”

Ref 24 showed this for both human fMRI and macaque cell-recording data.

(9) “Interestingly, mid-level features were the most useful in identifying object categories, suggesting the significant contributions of mid-level representations in accurate object identification.”

This sentence repeats the same point after “suggesting”.