Métodos de síntesis de ferrocenos homoquirales

Many theories of memory have been proposed that link hippocampal physiology to specific computations that could underlie learning and memory. Perhaps the most influential of these theories is the ‘Hebb-Marr’ model (Hebb, 1949, Marr, 1971) along with its variants (Hasselmo et al., 1996, McNaughton and Morris 1987, McNaughton and Nadel 1990, Treves and Rolls 1994; O’Reilly and McClelland, 1994). The common thread in these models is that the role of the hippocampus is to store representations of correlated inputs in independent networks which can be queried with partial input to reinstate the pattern of activity that was present during initial learning. This reinstatement constitutes the episodic memory.

Nearly all computational theories of memory model learning as changes in the connection strength between pre- and post-synaptic neurons that are coactive during learning. This concept of memory was famously proposed by Donald Hebb (1949) and commonly summarized simply as: cells that fire together wire together. Anti-correlation between pre- and post-synaptic cell firing can also lead to decreases in the connection strength between cells (Markram et al., 1997), thus bidirectional changes in synaptic strength reflect patterns of coactivity.

Studies on activity-dependent changes in synaptic strength led to the hypothesis that associative learning can be accomplished by finding optimal connection strengths between every pre-synaptic cell and every post-synaptic cell. This set of optimal

connections strengths constitutes a weight matrix that defines the desired transform of the

input patterns to the associated output patterns (Rumelhart and McClelland, 1986;

McCulloch and Pitts, 1943; Oja, 1989; Hopfield, 1982; Anderson, 1977, Hinton 1989, Kohonen, 1984). Many learning rules for synaptic weight modification have been

proposed, though regardless of the precise strategy, weight matrices in general can define a transform between a set of inputs and a set of outputs.

However, when multiple items are sequentially stored within the same weight space, the adjustments to the synaptic weights driven by new items disrupts the ability of artificial neuron networks to recall old items. In linear systems this interference is proportional to the correlation of the stored input patterns (Frean and Robbin, 1999), however, in non-linear training regimes (e.g. feedforward backpropogation), new learning can severely disrupt stored memories in what has been termed ‘catastrophic interference’ (McCloskey and Cohen, 1989). The reason for this is that the solution to ideal weight matrix to associate two patterns of activity is distributed across all of the individual weights in the matrix. Big differences in retrieval accuracy can occur with two similar conformations of the synaptic weight matrix, highlighting the existence of ‘weight cliffs’ in the solution space for an ideal transform function between inputs and outputs (French, 1999).

Since interference is particularly detrimental for correlated patterns, one solution to the problem is to decorrelate the input patterns prior to storage (French, 1992).

Decorrelation can occur by mapping a smaller representational space onto a larger one (Marr, 1971) and also by representing inputs with a sparse network of active cells with the majority silent (French, 1999). The large number of mostly silent cells in the dentate

gyrus suggested that the first stop in the trisynaptic loop involves pattern separation of correlated inputs (McNaughton and Morris, 1987; O’Reilly and McClelland, 1994;

Hasselmo and Wyble, 1997). Adult neurogenesis in the dentate offers another

mechanism by which to map inputs onto virgin weight space, thereby ensuring minimal interference (Wiskott et al., 2006). The sparse dentate code is hypothesized to be the result of a ‘winner-take-all’ computation in which a few cells fire only when specific conjunctions of stimuli (e.g. item in location) are present and once active inhibit their neighbors (Hasselmo and Wyble, 1997; Sloviter and Brisman et al., 1995). The connections between the EC to the DG and from DG to CA3 are considered random (O’Reilly and McClelland, 1994) and therefore the EC input is mapped onto a few random dentate cells. The combination of a random graph and the DG pattern separation strategies ensures that correlated inputs will be orthogonalized within the hippocampus.

The pattern separation function of the dentate is thought to be the critical feature that allows storage of correlated episodic memories without catastrophic interference.

After the dentate, information is propagated to CA3, a region that is known to have axons which synapse widely back over CA3 to form an autoassociative or recurrent network. Once a pattern of activity is stored within the weight matrix of a recurrent network, subsequent presentations of degraded inputs can associatively recall the originally stored pattern, a process known as pattern completion (Hopfield, 1982;

Kohonen, 1972; Anderson, 1972). Therefore CA3 is hypothesized to be important for pattern completion, or recall, of familiar memories with degraded inputs (O’Reilly and McClelland, 1994) either from DG of from direction projections from EC.

Finally, CA3 inputs onto CA1 and finally back to the ECII. Hippocampal signals must be inverted back into a form that can be read by the EC. The CA1 to EC ‘point-to-point’ topography is considered evidence for this translator role and therefore CA1 patterns should reflect some of the correlations of the input structure, but, due to pattern separation in the DG and CA, represent correlated inputs with relatively distinct patterns (O’Reilly and McClelland, 1994; Hasselmo and Wyble, 1997).

In Hebb-Marr models, the pattern completion function of CA3 is thought to be limited to recreating a pattern of activity during encoding. Novel conjunctions of familiar elements would not associatively activate prior traces but instead create a new

representation that is ideally independent of those representing the previously

experienced conjunctions (Kumaran and McClelland, 2012; O’Reilly and McClelland, 1994). These transforms on the signal within the hippocampus, in the words and spirit of McClelland, “by nature tend to support episodic memory at the expense of capturing the higher order structure of a set of experiences” (Kumaran and McClelland, 2012). It is this final conclusion that I will argue against throughout the rest of this thesis.

1.4 The hippocampus supports memory schemas that define the relationships