IEEE Technology and Society Magazine - December 2015 - 44

behalf in the future. This can have
numerous potential benefits, such
as having smart contract-invoked
advocacy and services for yourself
in the case of your aging and incapacitation while still living. In the
case of digital mindfile uploads,
smart contracts on the blockchain
are exactly the kind of future thirdparty advocate that can verify and
exercise control over the physical
parameters of your reality, of your
existence as a digital intelligence.
You would enact smart contracts on
the blockchain to periodically confirm your run-time parameters and
conduct decentralized back-ups.
In an advanced society of billions of digital intelligences living and thriving in smart network
systems, there would need to be
sophisticated oracles (information
arbiters) accessed by blockchain
smart contracts; oracles as a service, a platform, or as a public good
[21]. The Wikipedia of the future
could include blockchain-based oracle services to look up the current
standard for digital mindfile processing, storage, and security as these
standards would likely be advancing over time. "You are running on
the current standard, Windows 36,"
your smart contract informs you.
The endgame of blockchain thinking
is these kinds of futuristic mechanisms that could move society
forward. Dynamic oracle services
accessible by smart contracts on
universal public blockchains could
help to create a system of checks
and balances where digital intelligences could feel comfortable not
only in their basic survival, but also
in their future growth potential.

Blockchain Thinking:
The Architectural Proposal
Now having seen the vision for how
blockchain thinking might unfold,
the next step is articulating more
specifically it might work. Conceiv-

AI Research: Deep Mind's
Neural Turing Machines

ing of blockchains as universal
transaction systems is useful in the
case of blockchain thinking, particularly the built-in tracking and
account ability fea ture. Ever y
transaction can be recorded and
reviewed on-demand by any person,
machine code, or smart contract
with access to the system at any
later date. Three areas in the blockchain thinking architecture of inputprocessing-output are outlined:
memory, storage, and file-serving;
processing; and utility functions and
output. This is necessarily a gross
oversimplification of both biological
thinking and machine intelligence,
however an attempt to set forth a
basic approach.

In AI research, memory may start
to be seen as a standard discrete
external component. This is the
approach taken by Deep Mind, a
London-based startup purchased
by Google in 2014. Deep Mind has
created what it calls Neural Turing
Machines that attempt to mimic the
short-term memory of the human
brain by linking neural networks to
external memory modules [6]. The
fundamental process of computing as classically conceived uses
an external memory which can be
written to and read from during the
course of a computation, but memory had not been previously included in neural net AI architectures.
In Turing's famous description of a
computer, the memory is the tickertape that passes back and forth
through the computer and which
stores symbols of various kinds for
later processing. Deep Mind has
incorporated this idea and extended the capabilities of neural networks by coupling them to external
memory resources. In this architecture, a neural network can store
variables in its memory and come
back to them later to use in a calculation. The Neural Turing Machine
learns like a conventional neural
network using the inputs it receives
from the external world, and it also
learns how to store this information
and when to retrieve it [13].

Input: Memory, Storage,
File-Serving
Perhaps the first and most straightforward element needed for thinking is memory. Without having to
address complicated and important
questions like "What is memory?"
or "Where and how are human
memories stored?" and "Do all
human memories have the same
profile or characterization?," for
blockchain computational purposes, a position can be articulated
that each memory is a discrete unit
and that these discrete units are
encoded and stored somewhere.
Blended or overlapping memories could be stored as separate
discrete units. The key move for
blockchain-enabled thinking is that
instead of having just one instance
of a memory, there could be arbitrarily many copies of a memory,
just as there can be many copies
of any digital file. Further that the
number and location of any stored
items, in this case memories, could
be optimized dynamically for system operations. In fact researchers
in both AI and neuroscience have
been moving towards a modular
managed approach with memory.
IEEE Technology and Society Magazine

Neuroscience Research:
Active Memory Management
and Blocking
Recent progress in "wetware"
neuroscience also encourages a
modular computational approach
to memory, where it may be possible to both enhance and block
memories [26]. Several substances
have been shown to enhance memory performance such as donepezil
(which improves memory retention);
∕

december 2015

Table of Contents for the Digital Edition of IEEE Technology and Society Magazine - December 2015