Contrastive Learning Treats the Other Examples as Part of the Question

Place two views of the same image in a batch, then change only the other images. The positive pair stays the same, but its training loss changes. In SimCLR, the model must identify a matching view among the alternatives present in that batch. Those alternatives are part of the question the model is asked to answer.

The SimCLR paper constructs two independently augmented views of each of N source images. An encoder maps each view to a representation; a projection head maps that representation into the space where the contrastive loss is applied. For one view, called the anchor, its sibling view is the positive. The other 2(N - 1) views, produced from the remaining source images, serve as negatives. The paper does not need a separately sampled pool or memory bank to provide them. It applies the loss in both directions for every pair, so each view takes a turn as an anchor.

Imagine an AI developer preparing visual embeddings for product photos. One batch contains a blue chair, a lamp, a bicycle, and a sofa, each with two augmented views. The chair view must score its sibling higher than the six views of the other three products. In another batch, replace the lamp and bicycle with two other chairs. The same blue-chair positive pair now faces visually closer alternatives. The training signal is harder, and the model is pressed to preserve distinctions that the first batch barely tested. This is a conceptual example of the loss, not an experiment reported in the paper.

For each anchor, SimCLR computes cosine similarity between its projected vector and every other projected vector. It divides each similarity by a temperature value, exponentiates it, and normalizes across every candidate except the anchor itself. The loss is the negative log probability assigned to the positive view. Because the denominator includes the positive and all in-batch negatives, a new or more similar negative can raise the loss even when the anchor and positive vectors have not changed. That is the precise sense in which batch composition changes the optimization task.

Temperature controls how sharply similarities compete. At a lower positive temperature, a small similarity difference produces a larger ratio between exponentiated scores, so high-scoring alternatives take more of the normalized weight. A higher temperature softens that competition. The paper reports that normalization and an appropriate temperature improved its linear evaluation results, and explains that the cross-entropy form weights negatives according to their relative difficulty. Treat temperature as a parameter to validate with the chosen batch construction, augmentations, and representation task, rather than assume a setting transfers unchanged. Changing it changes the effective emphasis within the current candidate set.

The augmentations also define what counts as the same example. SimCLR uses random crops, color distortions, and blur to make the two views. If both crops preserve the object, successful matching can encourage an embedding that survives those transformations. If a crop removes the only identifying feature, the supposed positive pair can become ambiguous. That second case is a design risk inferred from how the objective is built. The paper shows that composition of augmentations strongly affected the quality of learned image representations in its experiments; it does not imply that any stronger transformation is automatically better for a different dataset.

Increasing the batch size supplies more negatives per anchor. The paper studied larger batches and longer training and found that they benefited its setup. Still, “more negatives” also means more opportunities for another source image to represent the same semantic category. SimCLR’s rule treats views from different source images as negatives, even if a human would call both images chairs. That is a consequence of instance-based pairing, and its cost depends on the downstream task. If the desired embedding should group near-duplicate products or multiple photos of one item, a batch containing those instances can push them apart. A practitioner should examine duplicates and category overlap rather than assume that every other record is a useful contrast.

A further implementation detail matters for interpreting results. The paper applies contrastive loss after a learned nonlinear projection head but evaluates the encoder representation before that head. The projection can absorb information useful for the training objective while the earlier representation remains useful for later classification. Copying only the loss formula and evaluating the projected vectors as if they were the reported embeddings changes the method being compared.

For an embedding system used by an AI search or retrieval feature, design the batch and augmentation policy together. Identify what should remain equivalent across views, which distinct records may share meaning, and whether batches regularly contain genuinely challenging alternatives. Record batch size and temperature with the model configuration, because they alter the comparisons that train the encoder. Then judge the frozen encoder on a held-out retrieval or classification task reflecting the intended use. A low training loss only says that positive views were identifiable among the alternatives supplied during training.