Reading guide: Similarity, Interactive Activation, and Mapping

There's a lot here too. The overall key points to make sure you extract are the description of the human similarity task in the first part of the paper, and then how the SIAM model (described starting page 12) works on the task.

Background

Early cognitive models of similarity suggested that people essentially just compare the features of the two things and see how much overlap there is. Goldstone is setting out to show that this approach is too simplistic, and that the structural relationships between the features must be taken into account. He did a set of experiments, and created a computer model. (I made an online version of his model, which you can access here.)

Things to watch out for

• After the intro, the paper describes two previous models of similarity. There are some formulas (formulae?) in here, but you don't need to be able to use them. It would be good if you can understand the high-level view of what they claim about human simmilarity judgments.
• In the Similarity and Alignment section, Goldstone lays out the general idea of how structure comes into similarity judgments. This sets up the rest of the paper.
• Next, Goldstone describes two experiments that he did to explore how structure affects human similarity judgments. Key concepts: MIPS and MOPS. (These relate to alignment in the previous section.)
• At the end of p. 11, the description of the SIAM model begins. This is a connectionist model (although unlike most other connectionist models, this one does not learn). Connectionist models are based loosely on the architecture of the brain, where nodes (neurons) are connected to other nodes by connections or arcs (synapses). The nodes essentially collect the incoming signals, and give a proportional output signal. (i.e. the bigger the input, the bigger the output.) Each connection has a "weight" that is multiplied by the output value of the node that it comes from, and this provides the input value of that connection to the to-node. The weight can be positive or negative. So if the from-node has a big positive output value, and the connection has a big negative weight, the input from this connection to the to-node is a big negative number. If the weight on the connection is near zero, the effect of the from-node on the to-node is minimized.
• In SIAM, each node represents a correspondence between two objects, features, or relationships in the two scenes. For example, one node will represent the strength of the hypothesis that Object A in one scene best corresponds with Object C in the other scene. If this node ends up with a high activation/value, then the model thinks that these two objects really do correspond.
• How would it end up with a high value? Features of the objects between the scenes can either match or not match. If they match, the corresponding nodes in the network get a high value input. The network goes through cycles by "spreading activition", i.e. each nodes outputs are multiplied by the weights of the outgoing connections, and then the resulting activations added into the activations of the to-nodes. Run the network for a few cycles, and it should come up with a determination of what the correspondences between the objects are. Add up the activations of the feature nodes, and you get the overall similarity between the scenes.
• In the rest of this section, Goldstone compares SIAM to other models of similarity.
• In the rest of the paper, Goldstone presents other Experiments to test out his model. Just try to get the gist of these, then pick up the conclusions at the end.