Understanding, Interpreting and Learning Representations in Deep Neural Networks
Bibliographic record
Abstract
<p>Deep Neural Networks (DNNs) have achieved state-of-the-art results in many computer vision tasks; however, DNNs have faced criticism for their lack of interpretability. Given the pervasiveness of DNNs in a multitude of applications, it is of paramount importance to fully understand the internal representations and behaviour of DNNs since safe and comprehensible utilization of DNN models is required before incorporating them into decision making processes for real-world applications. In this dissertation, we present several contributions towards understanding, interpreting, and learning representations in DNNs with an emphasis on studying absolute position information, interpreting latent representations to estimate certain semantic concepts, and learning robust representation. First, we study how much absolute position information is encoded in Convolutional Neural Networks (CNNs) as well as the source of this absolute position information. Our experiments reveal that a surprising degree of absolute position information is encoded in commonly used CNNs and zero padding enables CNNs to encode position information. Next, we analyze the relationship between boundary effects and padding in CNNs with respect to absolute position information. We also demonstrate how a CNN contains positional information in the latent representations if there exists a global pooling layer in the forward pass. We demonstrate that absolute position information is encoded based on the ordering of the channel dimensions, while semantic information is largely not. Second, we perform an empirical study on the ability of DNNs to encode shape information on a neuron-to-neuron and per-pixel level and show evidence that, while DNNs rely on texture information to recognize an object, a substantial amount of shape information is also encoded in DNNs. We further propose a new objective function for increasing a DNN’s ability to encode shape information by maximizing the mutual information between a network’s representations of two stylized images which share the same shape. Finally, we study the feature binding problem and present the first work which applies image blending to learn a robust representation for dense image labeling. Overall, we strongly believe the findings and demonstrated applications in this dissertation will benefit research areas concerned with understanding the different properties of DNNs.</p>
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
How this classification was reachedexpand
Full frame distilled prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. Learned from the 10,348 direct Codex labels and 10,348 direct Gemma labels. Candidate is the union of thresholded teacher heads; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels or direct frontier model labels.
Codex and Gemma teacher scores by category
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.000 | 0.000 |
| Meta-epidemiology (narrow) | 0.000 | 0.000 |
| Meta-epidemiology (broad) | 0.000 | 0.000 |
| Bibliometrics | 0.000 | 0.001 |
| Science and technology studies | 0.000 | 0.000 |
| Scholarly communication | 0.000 | 0.000 |
| Open science | 0.001 | 0.004 |
| Research integrity | 0.000 | 0.002 |
| Insufficient payload (model declined to judge) | 0.000 | 0.000 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from itClassification
machine, unvalidatedMachine predicted; a candidate call from one teacher head, not a consensus.
How this classification was reached, model by model and score by score, is at the end of the page under "How this classification was reached".