Tag: vision

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.

 

 

Do view-invariant brain representations of actions arise within 200 ms of viewing?

[R7I7]

Humans can rapidly visually recognise the actions people around them are engaged in and this ability is important for successful behaviour and social interaction. Isik et al. presented human subjects with 2-second video clips of humans performing actions while measuring brain activity with MEG. The clips comprised 5 actions (walk, run, jump, eat, drink) performed by each of five different actors and video-recorded from each of five different views (only frontal and profile used in MEG). Results show that action can be decoded from MEG signals arising about 200 ms after the onset of the video, with decoding accuracy peaking after about 500 ms and then decaying while the stimulus is still on, with a rebound after stimulus offset. Moreover, decoders generalise across actors and views. The authors conclude that the brain rapidly computes a representation that is invariant to view and actor.

ScreenShot738

Figure from the paper. Legend from the paper with my modifications in brackets: [Accuracy of action decoding (%) from MEG data as a function of time after video onset]. We can decode [which of five actions was being performed in the video clip] by training and testing on the same view (‘within-view’ condition), or, to test viewpoint invariance, training on one view (0 degrees [frontal, I think, but this should be clarified] or 90 degrees [profile]) and testing on the second view (‘across view’ condition). Results are each [sic] from the average of eight different subjects. Error bars represent standard deviation [across what?]. Horizontal line indicates chance decoding accuracy. […] Lines at the bottom of plot indicate significance with p<0.01 permutation test, with the thickness of the line indicating [for how many of the 8 subjects decoding was significant]. [Note the significant offset response after the 2-s video (whose duration should be indicated by a stimulus bar).]

 

The rapid view-invariant action decoding is really quite amazing. It would be good to see more detailed analyses to assess the nature of the signals enabling this decoding feat. Of course, 200 ms already allows for recurrent computations and the decodability peak is at 500 ms, so this is not strong evidence for a pure feedforward account.

The generalisation across actors is less surprising. This was a very controlled data set. Despite some variation in the appearance of the actors, it seems plausible that there would be some clustering of the vectors of pixels across space and time (or of features of a low-level visual representation) corresponding to different actors performing the same action seen from the same angle.

In separate experiments, the authors used static single frames taken from the videos and dynamic point-light figures as stimuli. These reduced form-only and motion-only stimuli were associated with diminished separation of actions in the human brain and in model representations, and with diminished human action recognition, suggesting that form and motion information are both essential to action recognition.

I’m wondering about the role of task-related priors. Subjects were performing an action recognition task on this controlled set of brief clips during MEG while freely viewing the clips (though this is not currently clearly stated). This task is likely to give rise to strong prior expectations about the stimulus (0 deg or 90 deg, one of five actions, known scale and positions of key features for action discrimination). Primed to attend to particular diagnostic features and to fixate in certain positions, the brain will configure itself for rapid dynamic discrimination among the five possible actions. The authors present a group-average analysis of eye movements, suggesting that these do not provide as much information about the actions as the MEG signal. However, the low-dimensional nature of the task space is in contrast to natural conditions, where a wider variety of actions can be observed and view, actor, size, and background vary more. The precise prior expectations might contribute to the rapid discriminability of the actions in the brain signals.

The authors model the results in the framework of feedforward processing in a Hubel-and-Wiesel/Poggio-style model that alternates convolution and max-pooling to simulate responses resembling simple and complex cells, respectively. This model is extended here to process video using spatiotemporal filter templates. The first layer uses Gabor filters, higher layers use templates in the first layer matching video clips in the stimulus set. The authors argue that this model supports invariant decoding and largely accounts for the MEG results.

Like the subjects, the model is set up to process the restricted stimulus space. The internal features of the model were constructed using representational fragments from samples from the same parametric space of videos. The exact videos used to test the models were not used for constructing the feature set. However, if I understood correctly, videos from the same restricted space (5 actions, 5 actors, 5 views) were used. Whether the model can be taken to explain (at a high level of abstraction) the computations performed by the brain depends critically on the degree to which the model is not specifically constructed for the (necessarily very limited) 5-action controlled stimulus space used in the study.

As the authors note, humans still outperform computer vision models at action recognition. How does the authors’ own model perform on less controlled action videos? If it the model cannot perform the task on real-world sensory input, can we be confident that it captures the way that the human brain performs the task? This is a concern in many studies and not trivial to address. However, the interpretation of the results should engage this issue.

 

Strengths

  • Controlled stimulus set: The set of video stimuli (5 actions x 5 actors x 5 views x 26 clips = 3250 2-sec clips) is impressive. Assembling this set is an important contribution to the field. The set is condition-rich (compared to typical stimulus sets used in cognitive neuroscience studies) and seems to strike a good balance between control and realism. This set could be a driver of progress if it were to be used in multiple modelling and empirical studies.
  • Combination of brain-activity measurements and a simple computational model, which provides a useful starting point for modelling the recognition of dynamic actions, as it is minimal and standard in many respects: a feedforward model in the HMAX framework, extended from spatial to spatiotemporal filters.

 

Weaknesses

  • Controlled stimulus set: The set of video stimuli is very restricted compared to real-world action recognition. For the brain data, this means that subjects might have rapidly formed priors about the stimuli, enabling them to configure their visual systems (e.g. attentional templates, fixation targets) for more rapid recognition of the 5 actions than is possible in real-world settings. This limitation is shared with many studies in our field and difficult to overcome without giving up control (which is a strength, see above). I therefore suggest addressing this problem in the discussion.
  • The model uses features based on spatiotemporal patterns sampled from the same restricted stimulus space. Although non-identical clips were used, the videos underlying the representational space appear to share a lot with the experimental stimuli (same 5 actions, same 5 views, same background?, same actors?). I would therefore not expect this model to work well on arbitrary real-world action video clips. This is in contrast to recent studies using deep convolutional neural nets (e.g. Khaligh-Razavi & Kriegeskorte 2014), where the models were trained without any information about the (necessarily restricted) brain-experimental stimulus set and can perform recognition under real-world conditions.
  • Only one model (in two variants) is tested. In order to learn about computational mechanism, it would be good to test more models.
  • MEG data were acquired during viewing of only 50 of the clips (5 actions x 5 actors x 2 views).
  • Missing inferential analyses: While the authors employ inferential analyses in single subjects and report number of significant subjects, few hypotheses of interest are formally statistically tested. The effects interpreted appear quite strong, so the results described above appear solid nevertheless (interpretational caveats notwithstanding).

 

Overall evaluation

This is an ambitious study describing results of a well-designed experiment using a stimulus set that is a major step forward. The results are an interesting and substantial contribution to the literature. However, the analyses could be much richer than they currently are and the interpretation of the results is not straightforward. Stimulus-set-induced priors may have affected both the neural processing measured and the model (which used templates from stimuli within the controlled video set). Results should be interpreted more cautiously in this context.

Although feedforward processing is an important part of the story, it is not the whole story. Recurrent signal flow is ubiquitous in the brain and essential to brain function. In engineering, similarly, recurrent neural networks are beginning to dominate spatiotemporal processing challenges such as speech and video recognition. The fact that the MEG data are presented as time courses, revealing a rich temporal structure, and the model analyses are bar graphs illustrates the key limitation of the model.

It would be great to extend the analyses to reveal a richer picture of the temporal dynamics. This should include an analysis of the extent to which each model layer can explain the representational geometry at each latency from stimulus onset.

 

Future directions

In revision or future studies, this line of work could be extended in a number of ways:

  • Use multiple models that can handle real-world action videos. The authors’ controlled video set is extremely useful for testing human and model representations, and for comparing humans to models. However, to be able to draw strong conclusions, the models, like the humans, would have to be trained to recognise human actions under real-world conditions (unrestricted natural video). In addition, it would be good to compare the biological representational dynamics to both feedforward and recurrent computational models.
  • To overcome the problem of stimulus-set related priors, which make it difficult to compare representational dynamics measured for restricted stimulus sets to real-world recognition in biological brains, one could present a large set of stimuli without ever presenting a stimulus twice to the same subject. Would the action category still be decodable at 150 ms with generalisation across views? Would a feedforward computer vision model trained on real-world action videos be able to predict the representational dynamics?
  • The MEG analyses could use source reconstruction to enable separate analyses of the representational dynamics in different brain regions.
  • It would be useful to have MEG data for the full stimulus set of 5 actions x 5 actors x 5 viewpoints = 125 conditions. The representational geometries could be analysed in detail to reveal which particular action pairs become discriminable when with what level of invariance.

 

 

Particular suggestions for improvements of this paper

(1) Present more detailed results

It would be good to see results separately for each pair of actions and each direction of crossdecoding (0 deg training -> 90 deg testing, and 90 deg training -> 0 deg testing). Regarding the former, eating and drinking involve very similar body postures and motions. Is this reflected in the discriminability of these actions?

Regarding, the decoding generalisation across views, you state:

“We decoded by training only on one view (0 degrees or 90 degrees), and testing on a second view (0 degrees or 90 degrees).”

Was the training set exclusively composed on 0 degree (frontal?) and the test set exclusively of 90 degree (side view?), and vice versa? In case the test set contained instances of both views (though of course, not for the same actor and action), results are more difficult to interpret.

 

(2) Discuss the caveats to the current interpretation of the results

Discuss the question whether priors resulting from subjects understanding of the restricted stimulus set might have affected the processing of the stimuli. Consider the involvement of recurrent computations within 200 ms and discuss the continuing rise of decodability until 500 ms. Discuss the possibility that the model will not generalise to action recognition in the wild.

 

(3) Test several control models

Can Gabor, HMAX, and deep convolutional neural net models support similarly invariant action decoding? These models are relatively easy to test, so I think it’s worth considering this for revision. Computer vision models trained on dynamic action recognition could be left to future studies.

 

(4) Test models by comparison of its representations with the brain representations

The computational model is currently only compared to the data at the very abstract level of decoding accuracy. Can the model predict the representations and representational dynamics in detail? It might be difficult to use the model to predict the measured channels. This would require the fitting of a linear model predicting the measured channels from the model units and the MEG data (acquired for only 5 actions x 5 actors x 2 views = 50 conditions) might be insufficient. However, representational dynamics could be investigated in the framework of representational similarity analysis (50 x 50 representational dissimilarity matrices) following Carlson et al. (2013) and Cichy et al. (2014). Note that this approach does not require fitting a prediction model and so appears applicable here. Either approach would reveal the dynamic prediction of the feedforward model (given dynamic inputs) and where its prediction diverges from the more complex and recurrent processes in the brain. This would promise to give us a richer and less purely confirmatory picture of the data and might show the merits and limitations of a feedforward account.

 

(5) Perform temporal cross-decoding

Temporal crossdecoding (Carlson et al. 2013, Cichy et al. 2014) could be used to more richly characterise the representational dynamics. This would reveal whether representations stabilise in certain time windows, or keep tumbling through the representational space even as stimuli are continuously decodable.

 

(6) Improve the inferential analyses

I don’t really understand the inference procedure in detail from the description in the methods section.

“We recorded the peak decoding accuracy for each time bin,…”

What is the peak decoding accuracy for each time bin? Is this the maximum accuracy across subjects for each time bin?

“…and used the null distribution of peak accuracies to select a threshold where decoding results performing above all points in the null distribution for the corresponding time point were deemed significant with P < 0.01 (1/100).”

I’m confused after reading this, because I don’t understand what is meant by “peak”.

The inference procedure for the decoding-accuracy time courses seems to lack formal multiple-testing correction across time points. Given enough subjects, inference could be performed with subject as a random effect. Alternatively, fixed-effects inference could be performed by permutation, averaging across subjects. Multiple testing across latencies should be formally corrected for. A simple way to do this is to relabel the experimental events once, compute an entire decoding time course, and record the peak decoding accuracy across time (or if this is what was done, it should be clearly described). Through repeated permutations, a null distribution of peak accuracies can be constructed and a threshold selected that is exceeded anywhere under H0 with only 5% probability, thus controlling the familywise error rate at 5%. This threshold could be shown as a line or as the upper edge of a transparent rectangle that partially obscures the insignificant part of the curve.

For each inferential analysis, please describe exactly what the null hypothesis was, what event-labels are exchangeable under this null hypothesis, and how the null distribution was computed. Also, explain how the permutation test interacted with the crossvalidation procedure. The crossvalidation should ideally generalise to new stimuli and label permutation be wrapped around this entire procedure.

“Decoding analysis was performed using cross validation, where the classifier was trained on a randomly selected subset of 80% of data for each stimulus and tested on the held out 20%, to assess the classifier’s decoding accuracy.”

Does this apply only to the within-view decoding? In the critical decoding analysis with generalisation across views, it cannot have been 20% of the data in the held-out set, since 0-deg views were used for training and 90-deg views for testing (and vice versa). If only 50% of the data were used for training there, why didn’t performance suffer given the smaller training set compared to the within-view decoding?

It would also be good to have estimates and inferential comparisons of the onset and peak latencies of the decoding time courses. Inference could be performed on a set of single-subject latency differences between two conditions modelling subject as a random effect.

 

(7) Qualify claims about biological fidelity of the model

The model is not really “designed to closely mimic the biology of the visual system”, rather its architecture is inspired by some of the features of the feedforward component of the visual hierarchy, such as local receptive fields of increasing size across a hierarchy of representations.

 

(8) Open stimuli and data

This work would be especially useful to the community if the video stimuli and the MEG data were made openly available. To fully interpret the findings, it would also be good to be able to view the movie clips online.

 

(9) Further clarify the title

The title “Fast, invariant representation for human action in the visual system” is somewhat unclear. What is meant are representations of perceived human actions, not representations for action. “Fast, invariant representation of visually perceived human actions” would be better, for example.

 

(10) Clarify what stimuli MEG data were acquired for

The abstract states “We use magnetoencephalography (MEG) decoding and a computational model to study action recognition from a novel dataset of well-controlled, naturalistic videos of five actions (run, walk, jump, eat, drink) performed by five actors at five viewpoints.” This suggests that MEG data were acquired for all these conditions. The methods section clarifies that MEG data were only recorded for 50 conditions (5 actions x 5 actors x 2 views). Here and in the legend of Fig. 1, it would be better to use the term “stimulus set” in place of “data set”.

 

(11) Clarify whether subjects were fixating or free viewing

Were subjects free viewing or fixating? This should be explicitly stated and the choice motivated in either case.

 

(12) Make figures more accessible

The figures are not optimal. Every panel showing decoding results should be clearly labelled to state what variables the crossvalidation procedure tested for generalisations across. For example, a label (in the figure itself!) could be: “decoding brain representations of actions with invariance to actor and view”. The reader shouldn’t have to search in the legend to find this essential information. Also every figure should have a stimulus bar depicting the period of stimulus presence. This is important especially to assess stimulus-offset-related effects, which appear to be present and significant.

Fig. 3 is great. I think it would be clearer to replace “space-time dot product” with “space-time convolution”.

 

(13) Clarify what the error bars represent

“Error bars represent standard deviation.”

Is this the standard deviation across the 8 subjects? Is it really the standard deviation or the standard error?

 

 (14) Clarify what we learn from the comparison between the structured and the unstructured model

For the unstructured model, won’t the machine learning classifier learn to select random combinations that tend to pool across different views of one action? This would render the resulting system computationally similar.

 

 

 

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”.