Jrcb4 The Impact of Artificial Intelligence on Learning final

Figure 2: Organization of a perceptron 
Source: Adapted from Rosenblatt, 1958
. 
The distinctive characteristics of most neural AI systems are their simple behaviouristic 
learning models, very high computational needs during learning, and their need for data. 
For these systems, the availability of data is the most critical success factor. Using 
27
Many excellent histories of AI and cognitive science exist that describe the interdependent development of 
computers, cognitive psychology, and artificial intelligence. See, e.g., McCorduck (1979), Gardner (1987) 
and Boden (2016). 
28
Rosenblatt (1958). Hebb was, in turn, influenced by Rashevsky’s work on neural networks. Hayek’s 
connectionist model of learning is described in Hayek (1952). 
29
Current deep-learning architectures use computational "backpropagation" of output error during learning to 
adjust network weights. In contrast, Rosenblatt's perceptron used feedback connections from its output 
layer for learning how to separate different input patterns. Although deep-learning networks are essentially 
perceptrons, Rosenblatt used in 1958 a vacuum tube computer that able to do about 12,000 mathematical 
additions or multiplications per second (12kFLOPS). Google's newest tensor processing compute ”pods,” 
announced in May 2018, can run more than hundred petaflops when training a machine learning system. 
That is 100,000,000,000,000,000 multiplications per second. This, in itself, is superhuman: If every person 
on Earth would make one multiplication every second, about ten million planet Earths would be needed to 
achieve the same computational capability. 
 

---

12 
biological terminology, they could be called “datavores.” Because of this, we call this the 
“data-based” approach to AI.
30