Understanding self-supervised Learning Dynamics without Contrastive Pairs

Abstract: While contrastive approaches of self-supervised learning (SSL) learn representations by minimizing the distance between two augmented views of the same data point (positive pairs) and maximizing views from different data points (negative pairs), recent \emph{non-contrastive} SSL (e.g., BYOL and SimSiam) show remarkable performance {\it without} negative pairs, with an extra learnable predictor and a stop-gradient operation. A fundamental question arises: why do these methods not collapse into trivial representations? We answer this question via a simple theoretical study and propose a novel approach, DirectPred, that \emph{directly} sets the linear predictor based on the statistics of its inputs, without gradient training. On ImageNet, it performs comparably with more complex two-layer non-linear predictors that employ BatchNorm and outperforms a linear predictor by $2.5\%$ in 300-epoch training (and $5\%$ in 60-epoch). DirectPred is motivated by our theoretical study of the nonlinear learning dynamics of non-contrastive SSL in simple linear networks. Our study yields conceptual insights into how non-contrastive SSL methods learn, how they avoid representational collapse, and how multiple factors, like predictor networks, stop-gradients, exponential moving averages, and weight decay all come into play. Our simple theory recapitulates the results of real-world ablation studies in both STL-10 and ImageNet. Code is released https://github.com/facebookresearch/luckmatters/tree/master/ssl.

• The paper provides a detailed analysis of gradient dynamics in self-supervised learning, revealing that non-traditional settings can enhance model performance.
• It introduces a mathematical framework for gradient updates in SimCLR using intra- and inter-augmentation covariance operators to manage data variance.
• The study demonstrates that tuning hyperparameters like λ and τ in the decoupled NCE loss leads to improved negative intra-augmentation covariance and overall SSL optimization.

An Evaluation of Gradient Dynamics in Self-Supervised Learning

Yuandong Tian's paper explores the intricacies of gradient dynamics in @@@@1@@@@ (SSL), specifically extending the analysis beyond the conventional balance condition, $$\frac{\partial L}{\partial r_+} + \frac{\partial L}{\partial r_-} = 0$$. This study addresses instances when this condition does not hold, positing that improved results can be achieved under these circumstances.

Gradient Dynamics in SimCLR

The paper introduces a detailed analysis of the SimCLR framework, highlighting the gradient update rule at a layer $l$ as:

$$\text{vec}(\Delta W_l) = OP_l\text{vec}(W_l) = (-\beta EV_l + VE_l)\text{vec}(W_l)$$

The operators $EV_l$ and $VE_l$ represent intra-augmentation and inter-augmentation covariance, respectively. The expression characterizes how variations within augmented data ($-\beta EV_l$) are managed to reduce covariate variance within data augmentation, influencing the learning dynamics positively.

Examination of Decoupled NCE Loss

The paper further explores the decoupled Noise Contrastive Estimation (NCE) loss, introducing the gradient updates with respect to $r_+$ and $r_{k-}$. By manipulating the terms:

$$^{\tau,\lambda} = r_+ + \lambda\log\left(e^{-r_+/\tau} + \sum_{k=1}^H e^{-r_{k-}/\tau}\right)$$

The paper derives:

$$\frac{\partial ^{\tau,\lambda}}{\partial r_+} + \sum_{k=1}^H \frac{\partial ^{\tau,\lambda}}{\partial r_{k-}} = 1 - \frac{\lambda}{\tau}$$

This demonstrates the impact of $\lambda$ and $\tau$ on the operator $OP_l$. The claim is made that adjusting $\lambda < \tau$ leads to a superior negative intra-augmentation covariance operator, corroborating findings in related literature that suggest improved performance under these conditions.

Implications and Future Directions

The implications of this research are both practical and theoretical. Practically, the findings suggest methods for optimizing SSL algorithms by considering conditions that deviate from traditional assumptions. Theoretically, it enhances the understanding of gradient dynamics and covariance operations within SSL frameworks. Future research could further explore the optimization of hyperparameters like $\lambda$ and $\tau$ across diverse SSL tasks and potentially extend these analyses to other SSL frameworks such as BYOL.

In conclusion, the paper provides a nuanced extension of the gradient dynamics analysis in SSL, offering insights that can augment the performance of SSL models under non-standard conditions. These contributions are of particular interest to researchers aiming to refine SSL methodologies for diverse applications.