Explicit information for category-orthogonal object properties increases along the ventral stream Hong et al., Nat. Neuro (2016). All figures are from the linked paper.

Introduction & Quick Summary

We’ve looked at multiple papers exploring and expanding the paradigm developed in Yamins et al (2014). Just as a recap, Yamins et al (2014)., implemented a convolutional neural network optimized for object classification task performance and was able to linearly decode macaque neural response that predicted about 48% of the explainable IT neural response variance (visit Brain-Score for more recent developments in predicting neural response of various areas along the visual stream). This CNN, which the paper calls hierarchical modular optimization (HMO) model, also performed object classification at the human performance level.

This paper probes whether ventral visual areas build a robust representation of ‘category-orthogonal’ object properties such as position, size, aspect ratio, etc. It also tackles whether these ‘category-orthogonal’ object property information is lost or preserved along the visual stream.

Previous research have shown that robust representations of objects are built along the ventral visual stream to support visual object categorization. However, it cannot be the case that visual perception only concerns itself with categorization as humans are able to estimate other object-related properties such as object position, size, aspect ratio and orientation. While these category-orthogonal properties are often considered to be ‘nuisance’ variables that are thrown away along the ventral visual stream, our ability to relatively precisely estimate those properties show that some of these information is preserved.

figure-1

Figure 1: Illustration of possible scenarios

Hong et al., catalogs four hypotheses about the pattern of information for category-orthogonal properties across the ventral visual stream:

  • H1. ‘local coding’. View tuned units are aggregated across identity-preserving transformations at each scale to produce partially view-invariant units, which are themselves aggregated to produce invariance at a larger scale. Trade-off between increasing receptive field size and {categorization ability, orthogonal task performance}.

  • H2. Non-categorical properties rely on intermediate visual features, analogous to ‘border-ownership cells’ that have been discovered in V2. Therefore, information for category orthogonal properties peak in the middle of the ventral stream.

  • H3. Information for low-level orthogonal properties is not lost along the ventral hierarchy, but is instead preserved because it may be behaviorally useful.

  • H4. Information increases for the category-orthogonal object tasks, similar to the increase in information for categorization task.

This manuscript identified the pattern of information by recording neural response in IT and V4 as well as testing simulated V1 neural response to a large image set of real-world objects with simultaneous variations in object position, size, pose and background scene. In doing so, the authors quantified the amount of explicitly available information in each ventral stream processing stage. Moreover, they assessed the dependence of information amount on image complexity and how the information is distributed across neural population.

Hong et al., found that

  • for all tasks in the high variation image set…
    • amount of explicitly available information progressively increased along the ventral stream, consistent with H4
    • task information is broadly distributed throughout the IT population
  • for all tasks in the low variation image set…
    • increase in information along the ventral stream attenuated
    • in some cases, the pattern of information reversed

Furthermore, convolutional neural network optimized for categorization task performance, was sufficient in accurately predicting patterns of information along the ventral stream across tasks and image variation levels. Thus, the authors claim these results are evidence that the perception of category-orthogonal properties, like object category perception, is constructed by the ventral visual hierarchy and that the neural network model provides an insight into the principles of the underlying hierarchy.

Comparing task representations across cortical areas

Some individual sites displayed high level of task selectivity. For example, the sites in Supp. Fig. 2a displayed high object category selectivity for different objects while the sites in Supp. Fig. 2b shows positional selectivity response. As visible in Fig 3a, the performance of single best sites from IT (blue) always outperformed the single best sites from V4 (green).

supp-figure-2

Supplementary Figure 2 {a,b}: Characterization of IT and V4 responses

The authors then compared visual property encoding at the neural population level. As visible in comparing the performance levels shown in Fig. 3a and Fig. 3b, neural population performance levels for representing a task were higher than single site performance levels. And following the trend in single site performance comparison, IT populations (Fig. 3b blue) always performed better than V4 populations(Fig. 3b green). Furthermore, both IT and V4 populations were better performing than V1 models (Fig. 3b grey) and pixel controls (Fig. 3b black).

figure-3

Figure 3: Comparison between ventral cortical areas of object property information encoding in high-variation stimuli

Neural consistency with human performance patterns

To bring human performance patterns into the picture, the manuscript estimated, across neural population and tasks, the number of neural sites required to reach parity with human performance levels. For all tasks, IT population reached human performance parity with at most 2000 sites, with a mean of about a third of those sites. V4 population needed several orders of magnitute more sites and V1 population needed several orders of magnitude more than V4 population – so much so that V1 reaching human performance parity is essentially intractable.

Distribution of Information across IT sites

Next, the authors characterized whether category-orthogonal properties are estimated by dedicated subpopulations in the IT or are jointly distributed over IT neurons in an overlapping manner. Previous studies showing modularity of face, body and place selective neurons suggest the former hypothesis while studies that demonstrate IT units tuned to multiple visual properties suggest the former.

To test these hypotheses, the paper took a two step approach:

  1. Quantify distribution of information across sites for each task
  2. Quantify information overlap between pairs of tasks

Two metrics, sparseness and imbalance, were used for the first step.

  • Sparseness measures the distribution of task-response correlation
  • Imbalance measures relative dominance of task-response correlated vs anticorrelated sites

Positive weights of the linear decoder for each of 266 IT sites was defined to be task-response correlation while negative weights was defined to be a task-response anticorrelation. Fig. 5a show that the distribution of task-response correlated units were sparse (At P=0.5 level, nearly half of the tasks had sparseness indistinguishable from the standard normal distribution). Not only was the information sparsely distributed across sites for most tasks, imbalance metric were also similar to an equivalently sized standard normal distribution at the P=0.5 level (Fig. 5b). These results suggest that other from few tasks that require relatively selective units, information for most category-orthogonal tasks was normally distributed.

figure-5

Figure 5: Distribution and overlap of IT cortex site contribution across tasks

Computational modeling

In this section, the authors explore whether the HCNN that was successful in predicting neural responses to categorization task can also accurately predict neural responses to category-orthogonal tasks. Similar to the HCNN model used in Yamins et al (2014), the convolutional neural network model implemented in this paper was optimized for image categorization task performance.

And in alignment with the results of that previous study, although the model never used neural response data during training, the network’s top hidden layer was predictive of IT neural response for category-orthogonal tasks. Furthermore, when the authors investigated category- and category-orthogonal task performances for each model layer, they found increasing performance level with each successive hidden layer. This is in direct accordance with the neural results they describe above. Plots for this comparison can be found in Fig. 6d and Fig. 6e.

These results are significant in that non-categorical tasks showed hierarchy even though the network model was never trained on these specific tasks. Furthermore it demonstrates the robustness and usefulness of the network paradigm described in Yamins et al (2014).

figure-6

Figure 6: Computational modeling results

Questions

  1. Fig. 5c show that face discrimination task required neither high-sparseness nor high-imbalance. Studies conjecture and prove the existence of face patches in IT cortex. If face patches were real, shouldn’t the face discrimination task require a high level of sparseness as only a few sites should be highly informative for the task? I can accept the low measure of imbalance but very low measure of sparseness seems contradictory to previous studies.