Literature
Display
← All papers

The Unreasonable Effectiveness of Recurrent Neural Networks

Andrej Karpathy

Published · Andrej Karpathy’s blog ↗

Contents

Overview

Karpathy demonstrates character-level language modeling with LSTMs trained on raw text. The examples show learned syntax and local structure; fluent-looking output does not establish factual knowledge, valid mathematics, or executable programs. [2]

Recurrent Neural Networks

An RNN applies the same learned update repeatedly, combining each input with a hidden state that carries history. This supports variable-length inputs and outputs and sequential processing of images. Stacking recurrent layers adds depth; the experiments use LSTM updates rather than the simple tanh cell shown in the code. [5] [6]

Five recurrent-network input and output configurations
Each rectangle is a vector and arrows represent functions (e.g. matrix multiply). Input vectors are in red, output vectors are in blue and green vectors hold the RNN's state (more on this soon). From left to right: (1) Vanilla mode of processing without RNN, from fixed-sized input to fixed-sized output (e.g. image classification). (2) Sequence output (e.g. image captioning takes an image and outputs a sentence of words). (3) Sequence input (e.g. sentiment analysis where a given sentence is classified as expressing positive or negative sentiment). (4) Sequence input and sequence output (e.g. Machine Translation: an RNN reads a sentence in English and then outputs a sentence in French). (5) Synced sequence input and output (e.g. video classification where we wish to label each frame of the video). Notice that in every case are no pre-specified constraints on the lengths sequences because the recurrent transformation (green) is fixed and can be applied as many times as we like.
Animated glimpse-based recognition of handwritten digits
Animated recurrent generation of handwritten digits
rnn = RNN()
y = rnn.step(x) # x is an input vector, y is the RNN's output vector
class RNN:
  # ...
  def step(self, x):
    # update the hidden state
    self.h = np.tanh(np.dot(self.W_hh, self.h) + np.dot(self.W_xh, x))
    # compute the output vector
    y = np.dot(self.W_hy, self.h)
    return y
y1 = rnn1.step(x)
y = rnn2.step(y1)

Character-Level Language Models

Predict the next character from preceding characters, train with cross-entropy and backpropagation through the sequence, then sample a character and feed it back. One-hot inputs alone cannot distinguish repeated letters with different continuations; the hidden state must track context. Sampling temperature trades variety against predictable continuations. [8] [13]

Character-level recurrent network predicting the next letters in hello

Fun with RNNs

The following examples keep the same learning objective while changing the training corpus. Larger networks and more training improve surface consistency, but failures often reveal dependencies that exceed the learned memory.

Paul Graham generator

A model trained on essays reproduces paragraph structure and an essay-like voice. Its locally plausible sentences can drift in meaning, illustrating the gap between matching text statistics and sustaining an argument. [9] [10]

Shakespeare

Training on Shakespeare yields recognizable character labels and dramatic formatting. The generated dialogue imitates style without reliably maintaining a coherent scene. [11]

PANDARUS:
Alas, I think he shall be come approached and the day
When little srain would be attain'd into being never fed,
And who is but a chain and subjects of his death,
I should not sleep.

Second Senator:
They are away this miseries, produced upon my soul,
Breaking and strongly should be buried, when I perish
The earth and thoughts of many states.

DUKE VINCENTIO:
Well, your wit is in the care of side and that.

Second Lord:
They would be ruled after this chamber, and
my fair nues begun out of the fact, to be conveyed,
Whose noble souls I'll have the heart of the wars.

Clown:
Come, sir, I will make did behold your worship.

VIOLA:
I'll drink it.
VIOLA:
Why, Salisbury must find his flesh and thought
That which I am not aps, not a man and in fire,
To show the reining of the raven and the wars
To grace my hand reproach within, and not a fair are hand,
That Caesar and my goodly father's world;
When I was heaven of presence and our fleets,
We spare with hours, but cut thy council I am great,
Murdered and by thy master's ready there
My power to give thee but so much as hell:
Some service in the noble bondman here,
Would show him to her wine.

KING LEAR:
O, if you were a feeble sight, the courtesy of your law,
Your sight and several breath, will wear the gods
With his heads, and my hands are wonder'd at the deeds,
So drop upon your lordship's head, and your opinion
Shall be against your honour.

Wikipedia

Wikipedia training produces markup, headings, lists, and invented encyclopedia-like passages. These patterns emerge from next-character prediction; generated assertions and references should not be treated as facts. [12] [14]

Naturalism and decision for the majority of Arab countries' capitalide was grounded
by the Irish language by [[John Clair]], [[An Imperial Japanese Revolt]], associated 
with Guangzham's sovereignty. His generals were the powerful ruler of the Portugal 
in the [[Protestant Immineners]], which could be said to be directly in Cantonese 
Communication, which followed a ceremony and set inspired prison, training. The 
emperor travelled back to [[Antioch, Perth, October 25|21]] to note, the Kingdom 
of Costa Rica, unsuccessful fashioned the [[Thrales]], [[Cynth's Dajoard]], known 
in western [[Scotland]], near Italy to the conquest of India with the conflict. 
Copyright was the succession of independence in the slop of Syrian influence that 
was a famous German movement based on a more popular servicious, non-doctrinal 
and sexual power post. Many governments recognize the military housing of the 
[[Civil Liberalization and Infantry Resolution 265 National Party in Hungary]], 
that is sympathetic to be to the [[Punjab Resolution]]
(PJS)[http://www.humah.yahoo.com/guardian.
cfm/7754800786d17551963s89.htm Official economics Adjoint for the Nazism, Montgomery 
was swear to advance to the resources for those Socialism's rule, 
was starting to signing a major tripad of aid exile.]]
{ { cite journal | id=Cerling Nonforest Department|format=Newlymeslated|none } }
''www.e-complete''.

'''See also''': [[List of ethical consent processing]]

== See also ==
*[[Iender dome of the ED]]
*[[Anti-autism]]

===[[Religion|Religion]]===
*[[French Writings]]
*[[Maria]]
*[[Revelation]]
*[[Mount Agamul]]

== External links==
* [http://www.biblegateway.nih.gov/entrepre/ Website of the World Festival. The labour of India-county defeats at the Ripper of California Road.]

==External links==
* [http://www.romanology.com/ Constitution of the Netherlands and Hispanic Competition for Bilabial and Commonwealth Industry (Republican Constitution of the Extent of the Netherlands)]

<page>
  <title>Antichrist</title>
  <id>865</id>
  <revision>
    <id>15900676</id>
    <timestamp>2002-08-03T18:14:12Z</timestamp>
    <contributor>
      <username>Paris</username>
      <id>23</id>
    </contributor>
    <minor />
    <comment>Automated conversion</comment>
    <text xml:space="preserve">#REDIRECT [[Christianity]]</text>
  </revision>
</page>

Algebraic Geometry (Latex)

LaTeX samples resemble mathematical exposition, complete with theorem environments and diagrams. Mismatched environments and meaningless arguments reveal limitations in long-range consistency and mathematical validity. [16] [17]

Generated algebraic-geometry text: first pair of LaTeX pages
Generated algebraic-geometry text: second pair of LaTeX pages
\begin{proof}
We may assume that $\mathcal{I}$ is an abelian sheaf on $\mathcal{C}$.
\item Given a morphism $\Delta : \mathcal{F} \to \mathcal{I}$
is an injective and let $\mathfrak q$ be an abelian sheaf on $X$.
Let $\mathcal{F}$ be a fibered complex. Let $\mathcal{F}$ be a category.
\begin{enumerate}
\item \hyperref[setain-construction-phantom]{Lemma}
\label{lemma-characterize-quasi-finite}
Let $\mathcal{F}$ be an abelian quasi-coherent sheaf on $\mathcal{C}$.
Let $\mathcal{F}$ be a coherent $\mathcal{O}_X$-module. Then
$\mathcal{F}$ is an abelian catenary over $\mathcal{C}$.
\item The following are equivalent
\begin{enumerate}
\item $\mathcal{F}$ is an $\mathcal{O}_X$-module.
\end{lemma}

Linux Source Code

A three-layer LSTM trained on concatenated Linux sources learns indentation, comments, pointer syntax, and license-like headers. Undefined variables, inconsistent return types, and vacuous conditions show why plausible source code need not compile or compute anything useful. [18] [19]

/*
 * Increment the size file of the new incorrect UI_FILTER group information
 * of the size generatively.
 */
static int indicate_policy(void)
{
  int error;
  if (fd == MARN_EPT) {
    /*
     * The kernel blank will coeld it to userspace.
     */
    if (ss->segment < mem_total)
      unblock_graph_and_set_blocked();
    else
      ret = 1;
    goto bail;
  }
  segaddr = in_SB(in.addr);
  selector = seg / 16;
  setup_works = true;
  for (i = 0; i < blocks; i++) {
    seq = buf[i++];
    bpf = bd->bd.next + i * search;
    if (fd) {
      current = blocked;
    }
  }
  rw->name = "Getjbbregs";
  bprm_self_clearl(&iv->version);
  regs->new = blocks[(BPF_STATS << info->historidac)] | PFMR_CLOBATHINC_SECONDS << 12;
  return segtable;
}
/*
 * If this error is set, we will need anything right after that BSD.
 */
static void action_new_function(struct s_stat_info *wb)
{
  unsigned long flags;
  int lel_idx_bit = e->edd, *sys & ~((unsigned long) *FIRST_COMPAT);
  buf[0] = 0xFFFFFFFF & (bit << 4);
  min(inc, slist->bytes);
  printk(KERN_WARNING "Memory allocated %02x/%02x, "
    "original MLL instead\n"),
    min(min(multi_run - s->len, max) * num_data_in),
    frame_pos, sz + first_seg);
  div_u64_w(val, inb_p);
  spin_unlock(&disk->queue_lock);
  mutex_unlock(&s->sock->mutex);
  mutex_unlock(&func->mutex);
  return disassemble(info->pending_bh);
}

static void num_serial_settings(struct tty_struct *tty)
{
  if (tty == tty)
    disable_single_st_p(dev);
  pci_disable_spool(port);
  return 0;
}

static void do_command(struct seq_file *m, void *v)
{
  int column = 32 << (cmd[2] & 0x80);
  if (state)
    cmd = (int)(int_state ^ (in_8(&ch->ch_flags) & Cmd) ? 2 : 1);
  else
    seq = 1;
  for (i = 0; i < 16; i++) {
    if (k & (1 << 1))
      pipe = (in_use & UMXTHREAD_UNCCA) +
        ((count & 0x00000000fffffff8) & 0x000000f) << 8;
    if (count == 0)
      sub(pid, ppc_md.kexec_handle, 0x20000000);
    pipe_set_bytes(i, 0);
  }
  /* Free our user pages pointer to place camera if all dash */
  subsystem_info = &of_changes[PAGE_SIZE];
  rek_controls(offset, idx, &soffset);
  /* Now we want to deliberately put it to device */
  control_check_polarity(&context, val, 0);
  for (i = 0; i < COUNTER; i++)
    seq_puts(s, "policy ");
}
/*
 *  Copyright (c) 2006-2010, Intel Mobile Communications.  All rights reserved.
 *
 *   This program is free software; you can redistribute it and/or modify it
 * under the terms of the GNU General Public License version 2 as published by
 * the Free Software Foundation.
 *
 *        This program is distributed in the hope that it will be useful,
 * but WITHOUT ANY WARRANTY; without even the implied warranty of
 *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
 *
 *  GNU General Public License for more details.
 *
 *   You should have received a copy of the GNU General Public License
 *    along with this program; if not, write to the Free Software Foundation,
 *  Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
 */

#include <linux/kexec.h>
#include <linux/errno.h>
#include <linux/io.h>
#include <linux/platform_device.h>
#include <linux/multi.h>
#include <linux/ckevent.h>

#include <asm/io.h>
#include <asm/prom.h>
#include <asm/e820.h>
#include <asm/system_info.h>
#include <asm/setew.h>
#include <asm/pgproto.h>

#define REG_PG    vesa_slot_addr_pack
#define PFM_NOCOMP  AFSR(0, load)
#define STACK_DDR(type)     (func)

#define SWAP_ALLOCATE(nr)     (e)
#define emulate_sigs()  arch_get_unaligned_child()
#define access_rw(TST)  asm volatile("movd %%esp, %0, %3" : : "r" (0));   \
  if (__type & DO_READ)

static void stat_PC_SEC __read_mostly offsetof(struct seq_argsqueue, \
          pC>[1]);

static void
os_prefix(unsigned long sys)
{
#ifdef CONFIG_PREEMPT
  PUT_PARAM_RAID(2, sel) = get_state_state();
  set_pid_sum((unsigned long)state, current_state_str(),
           (unsigned long)-1->lr_full; low;
}

Generating Baby Names

Training on a list of names yields new strings with name-like spelling patterns. Most displayed samples are absent from the training list, demonstrating recombination of learned character statistics. [20] [21]

Understanding what’s going on

Training snapshots and hidden-state visualizations offer two complementary views: which regularities appear first, and which contextual signals individual units may track.

The evolution of samples while training

Early samples acquire spaces and punctuation before common short words, longer words, and broader themes. The ordering suggests that local statistical constraints are learned before many long-range dependencies.

tyntd-iafhatawiaoihrdemot  lytdws  e ,tfti, astai f ogoh eoase rrranbyne 'nhthnee e 
plia tklrgd t o idoe ns,smtt   h ne etie h,hregtrs nigtike,aoaenns lng
"Tmont thithey" fomesscerliund
Keushey. Thom here
sheulke, anmerenith ol sivh I lalterthend Bleipile shuwy fil on aseterlome
coaniogennc Phe lism thond hon at. MeiDimorotion in ther thize."
we counter. He stutn co des. His stanted out one ofler that concossions and was 
to gearang reay Jotrets and with fre colt otf paitt thin wall. Which das stimn 
Aftair fall unsuch that the hall for Prince Velzonski's that me of
her hearly, and behs to so arwage fiving were to it beloge, pavu say falling misfort 
how, and Gogition is so overelical and ofter.
"Kite vouch!" he repeated by her
door. "But I would be done and quarts, feeling, then, son is people...."
"Why do what that day," replied Natasha, and wishing to himself the fact the
princess, Princess Mary was easier, fed in had oftened him.
Pierre aking his soul came to the packs and drove up his father-in-law women.

Visualizing the predictions and the “neuron” firings in the RNN

Some units correlate with URLs, bracketed markup, position within a span, or quotation state. These are suggestive visual interpretations of a distributed representation, rather than causal proofs that each unit implements the proposed rule. The figures retain both predicted characters and activation patterns. [22]

Character predictions and a hidden unit responding to URLs
Character predictions and a hidden unit responding to bracketed markup
Character predictions and a hidden unit tracking position inside bracketed markup
Character predictions and a hidden unit responding within the www prefix
Hidden-state visualizations: line position, markup, conditional statements, and less interpretable cells
Hidden-state visualizations: quotations, expression depth, and word-initial characters

Source Code

The accompanying char-rnn implementation trains and samples multilayer LSTMs; a smaller NumPy implementation exposes the core update. The discussion compares the flexibility and development workflow of the frameworks available in 2015. [2] [8] [23]

Further Reading

Applications span speech, translation, captioning, and visual attention. External-memory models decouple storage from computation, while soft attention makes addressing differentiable at the cost of reading every candidate location. Hard attention reduces that work but requires methods for learning through discrete choices, such as REINFORCE. [24] [25] [33] [34] [35] [36] [38] [39]

Conclusion

Character prediction can learn substantial structure from raw text with a small recurrent program. The examples motivate richer memory and attention mechanisms while making the limitations of surface imitation visible. Later links include a character n-gram comparison and readers’ experiments with recipes, speeches, and music. [54] [56] [59] [63]

I've the RNN with and works, but the computed with program of the 
RNN with and the computed of the RNN with with and the code
Concept demonstration

Open image ↗