Can BAM recall from either side?

Its defining design supports paired associations in both directions, although actual recall success depends on storage and input conditions.

Does every stable pair correspond to an intended memory?

No. Interference and model dynamics can produce unwanted outcomes.

Is BAM necessarily a deep neural network?

No. The classic model has two interacting layers. It is a historical associative-memory architecture, even when grouped with broader neural-network topics in a glossary.

AI Glossary / Deep learning / Intermediate

Bidirectional associative memory

Definition

Bidirectional associative memory, or BAM, is a neural associative-memory model that stores relationships between pairs of patterns and can use a cue on either side to help recall its partner. In the classic form, two groups of units connect to one another, and activity passes back and forth between them. The aim is paired-pattern recall, including recall from suitable imperfect cues, rather than exact lookup by a database address.

What it means

One pattern might encode a symbol and its partner a label. The two representations can have different lengths because each belongs to its own layer. Presenting the symbol can help retrieve the label, while presenting the label can help retrieve the symbol.

This differs from autoassociative recall, which reconstructs a pattern from a cue in the same pattern space. It also differs from bidirectional sequence models, where “bidirectional” commonly means processing a sequence from both ends. BAM is a specific feedback-memory architecture with its own storage and recall assumptions, not a general name for every model that uses information in two directions.

How it works

A basic BAM stores paired representations in a weight matrix connecting the two layers. A common construction adds outer products of the paired patterns. During recall, multiplying a cue by the connection matrix and applying an activation rule produces activity in the other layer.

The reverse pass uses the corresponding transposed connections. Alternating passes can settle into a stable pair under the specified model and update conditions.

Several stored associations can interfere. Stability therefore does not mean the intended association has been retrieved. Capacity, noise tolerance, and convergence behavior depend on the particular formulation, encoded patterns, thresholds, and updating scheme.

Three practical examples

1. Paired-code demonstration: students store small symbol vectors together with label vectors and test recall in both directions, observing that the two layers need not have equal sizes.

2. Corrupted-cue study: researchers alter selected components of a stored input and investigate whether the corresponding output pattern can still be recalled.

3. Association-capacity experiment: a team adds increasingly similar pattern pairs to a documented BAM implementation and measures interference, mistaken recall, and stable unwanted outcomes rather than assuming every pair remains recoverable.

Think of it this way

Think of a set of two-sided vocabulary cards: either side can remind you of the paired side. BAM adds a feedback process that repeatedly checks one side through the other. The analogy is limited because a numerical memory superimposes associations in shared connections. It does not necessarily retain each pair as an independent, perfectly reversible card, and ambiguous cues can lead to the wrong association.

Worked example

Store one bipolar pair: A = (+1, −1, +1) and B = (+1, −1). Define an unscaled weight matrix as their outer product, with rows (1, −1), (−1, 1), and (1, −1).

Present A as a row vector. Multiplying by the matrix gives (3, −3), whose signs recover B. Passing B back through the transpose gives (2, −2, 2), whose signs recover A.

Now use the corrupted cue (+1, +1, +1). Its forward result is (1, −1), so the same sign rule still produces B. The backward pass then restores A.

This is a deliberately easy one-pair example with no zero-valued activation sums. It does not establish performance with multiple interfering pairs or specify how ties should be handled. Different conventions may scale the matrix without changing these particular sign results.

Illustrative case study

Fictional case: Pair Harbour, an educational technology team, builds a BAM visualization associating simple icons with short binary labels. Early demonstrations use distinct icons and work as expected. The team then introduces icons that differ in only a small number of components.

Some altered cues settle into the wrong stored pairing. Instead of hiding those cases, the team displays the initial cue, each alternating pass, and the final pair. Learners can separate successful correction from mistaken convergence. No deployment result is claimed. The exercise shows why paired-memory behavior must be evaluated with ambiguous examples, not only cues already identical to stored patterns.

Common misconceptions and mistakes

BAM is not an ordinary key-value store with guaranteed exact reversal. Shared weights can produce interference and ambiguous recall.

The two layers need not have the same number of units. What matters is that the connection matrix dimensions match the two representations.

“Bidirectional” does not imply a causal relationship between the represented objects. Nor does a stable state certify that the output is the intended partner. Finally, similar-looking two-layer diagrams can describe different models. A restricted Boltzmann machine, for example, has a different probabilistic learning interpretation; architectural resemblance alone does not make it interchangeable with classic BAM.

Business use

For businesses, BAM is mainly relevant as a foundation for understanding associative retrieval and as a specialist research model. A useful evaluation should begin with a clearly defined paired-representation task and a reason to prefer approximate cue-based recall over exact lookup or another retrieval method.

Measure the cost of incorrect associations as well as the rate of successful recall. Include unfamiliar inputs and cases where several stored pairs are similar. If a demonstration cleans up a noisy input, preserve the original so inferred reconstruction is not mistaken for observed fact. The appropriate comparison is the complete retrieval workflow, including encoding and error handling, rather than the elegance of the feedback mechanism alone.

Check your understanding

In the worked example, the A layer has three components and the B layer has two. Must one vector be padded to the other's length?

Answer: No. A three-by-two connection matrix relates them, and the transpose supports the reverse direction. Equal layer sizes are not required.

Related terms

Associative memory
Autoassociative memory
Hopfield Network
Recurrent neural network
heteroassociation
outer product
attractor
content-addressable memory

Learn next

Reproduce the one-pair calculation in both directions, then add a second pair and inspect cross-interference. State the activation and tie-handling rules explicitly before comparing clean, corrupted, and unfamiliar cues.

Sources

Primary sources: Kosko, “Bidirectional associative memories,” https://sipi.usc.edu/~kosko/BAM.pdf ; Barra and colleagues, “Thermodynamics of bidirectional associative memories,” https://arxiv.org/html/2211.09694v2 . The original PDF opened without extractable text; the later primary paper's model description and equations were read. The unscaled one-pair example is original and does not invoke the later paper's thermodynamic-limit results.

Previous: Bidirectional Next: Bidirectional language model