Threshold Logic Unit
A threshold logic unit (TLU) is a computing element that forms a weighted sum of its binary or numeric inputs and outputs 1 if that sum reaches a threshold, and 0 otherwise. It is the hard-limiting, step-function ancestor of the modern artificial neuron: the same weighted-sum structure, but with an all-or-nothing output and, in its classic form, a small set of weights that are either designed by hand or adjusted by a simple error-correction rule.
The unit's behavior is fully described by one inequality. With inputs x1 through xn, weights w1 through wn and threshold θ, the output is 1 when w1·x1 + … + wn·xn ≥ θ and 0 when it falls short. Choose weights of 1 and 1 with a threshold of 2 and the unit computes AND, since only the input pair (1, 1) reaches 2; keep the weights and lower the threshold to 1 and it computes OR; a single weight of −1 with a threshold of 0 computes NOT. Because AND, OR and NOT are enough to build any Boolean function, networks of TLUs are computationally complete in principle, which is the observation at the core of the McCulloch-Pitts paper of 1943 (a TLU with fixed weights is essentially their formal neuron), and the reason electrical engineers in the late 1950s and 1960s studied threshold logic as a serious alternative to gates built from AND, OR and NOT.
Geometry explains both the power and the limit. The equation w·x = θ defines a straight line in two inputs, a plane in three and a hyperplane beyond; the TLU answers 1 on one side of it and 0 on the other, so it can compute exactly those Boolean functions whose true points can be separated from the false points by a single flat cut — the linearly separable functions. AND and OR pass this test. XOR (exclusive or), true when exactly one input is 1, does not: placing the four corners of the unit square on paper, the true corners (0,1) and (1,0) sit diagonally opposite each other, as do the false ones, and no single straight line can divide the two diagonals. Parity over more inputs, and many other natural functions, fail in the same way.
The TLU became the central object of the perceptron controversy. Frank Rosenblatt's perceptron (1958) is, in its simplest form, a TLU whose weights are trained by an error-driven rule, and his perceptron convergence theorem guarantees that the training procedure finds a separating set of weights in a finite number of corrections whenever the data are linearly separable. Marvin Minsky and Seymour Papert's book Perceptrons (MIT Press, 1969) then analyzed what single-layer threshold devices can and cannot compute, treating XOR and parity as emblematic and proving further results, such as limits on recognizing whether a figure is connected using units that each see only a small part of the image. The book's reputation often outran its contents: it did not prove that multilayer networks are useless, but it expressed doubt that a good learning method for them would be found, and, together with the funding climate of the era, it contributed to a long lull in neural-network research until backpropagation revived the field in the mid-1980s.
A frequent pitfall is to conflate the TLU's limitation with a limitation of neural networks as such. The hidden layer solves XOR in a few lines: one threshold unit detects OR, another detects AND, and an output unit fires when the first is on and the second is off. What a hand-built solution does not provide is a way to learn those hidden weights, and that missing training method, not the absence of expressive power, was the real obstacle of the 1960s. A second misreading runs the other way: since the step function has zero slope almost everywhere, gradient-based training cannot be applied to it directly, which is why later networks replaced the step with smooth or piecewise-linear activations, and why the TLU is now mostly a teaching model rather than a working component.
Its modern echoes are still concrete. The linear classifier at the heart of logistic regression and the support vector machine is a TLU in all but the final output rule, and the large-margin ideas behind the latter descend from perceptron analysis; binarized and quantized networks reintroduce hard thresholds to save memory and energy; and threshold gates remain a topic in circuit complexity, where constant-depth networks of them are a standard model. In teaching, the TLU is the cleanest demonstration that capability depends on representation: the same unit that cannot compute XOR alone computes it as part of a two-layer arrangement.
Read threshold logic unit as a straight-line decision with a hard yes or no at the edge: simple enough to analyze completely, limited by exactly the geometry of one cut, and important mostly for what it taught about stacking such cuts.